2 * Copyright © 2016 Red Hat.
3 * Copyright © 2016 Bas Nieuwenhuizen
5 * based in part on anv driver which is:
6 * Copyright © 2015 Intel Corporation
8 * Permission is hereby granted, free of charge, to any person obtaining a
9 * copy of this software and associated documentation files (the "Software"),
10 * to deal in the Software without restriction, including without limitation
11 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
12 * and/or sell copies of the Software, and to permit persons to whom the
13 * Software is furnished to do so, subject to the following conditions:
15 * The above copyright notice and this permission notice (including the next
16 * paragraph) shall be included in all copies or substantial portions of the
19 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
20 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
21 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
22 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
23 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
24 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
28 #include "util/mesa-sha1.h"
29 #include "util/u_atomic.h"
30 #include "radv_debug.h"
31 #include "radv_private.h"
32 #include "radv_shader.h"
34 #include "nir/nir_builder.h"
35 #include "spirv/nir_spirv.h"
37 #include <llvm-c/Core.h>
38 #include <llvm-c/TargetMachine.h>
42 #include "r600d_common.h"
43 #include "ac_binary.h"
44 #include "ac_llvm_util.h"
45 #include "ac_nir_to_llvm.h"
46 #include "vk_format.h"
47 #include "util/debug.h"
48 #include "ac_exp_param.h"
51 radv_pipeline_destroy(struct radv_device
*device
,
52 struct radv_pipeline
*pipeline
,
53 const VkAllocationCallbacks
* allocator
)
55 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; ++i
)
56 if (pipeline
->shaders
[i
])
57 radv_shader_variant_destroy(device
, pipeline
->shaders
[i
]);
59 if (pipeline
->gs_copy_shader
)
60 radv_shader_variant_destroy(device
, pipeline
->gs_copy_shader
);
62 vk_free2(&device
->alloc
, allocator
, pipeline
);
65 void radv_DestroyPipeline(
68 const VkAllocationCallbacks
* pAllocator
)
70 RADV_FROM_HANDLE(radv_device
, device
, _device
);
71 RADV_FROM_HANDLE(radv_pipeline
, pipeline
, _pipeline
);
76 radv_pipeline_destroy(device
, pipeline
, pAllocator
);
79 static void radv_dump_pipeline_stats(struct radv_device
*device
, struct radv_pipeline
*pipeline
)
81 unsigned lds_increment
= device
->physical_device
->rad_info
.chip_class
>= CIK
? 512 : 256;
82 struct radv_shader_variant
*var
;
83 struct ac_shader_config
*conf
;
86 unsigned max_simd_waves
;
87 unsigned lds_per_wave
= 0;
89 switch (device
->physical_device
->rad_info
.family
) {
90 /* These always have 8 waves: */
100 for (i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
101 if (!pipeline
->shaders
[i
])
103 var
= pipeline
->shaders
[i
];
107 if (i
== MESA_SHADER_FRAGMENT
) {
108 lds_per_wave
= conf
->lds_size
* lds_increment
+
109 align(var
->info
.fs
.num_interp
* 48, lds_increment
);
112 if (conf
->num_sgprs
) {
113 if (device
->physical_device
->rad_info
.chip_class
>= VI
)
114 max_simd_waves
= MIN2(max_simd_waves
, 800 / conf
->num_sgprs
);
116 max_simd_waves
= MIN2(max_simd_waves
, 512 / conf
->num_sgprs
);
120 max_simd_waves
= MIN2(max_simd_waves
, 256 / conf
->num_vgprs
);
122 /* LDS is 64KB per CU (4 SIMDs), divided into 16KB blocks per SIMD
126 max_simd_waves
= MIN2(max_simd_waves
, 16384 / lds_per_wave
);
128 fprintf(file
, "\n%s:\n",
129 radv_get_shader_name(var
, i
));
130 if (i
== MESA_SHADER_FRAGMENT
) {
131 fprintf(file
, "*** SHADER CONFIG ***\n"
132 "SPI_PS_INPUT_ADDR = 0x%04x\n"
133 "SPI_PS_INPUT_ENA = 0x%04x\n",
134 conf
->spi_ps_input_addr
, conf
->spi_ps_input_ena
);
136 fprintf(file
, "*** SHADER STATS ***\n"
139 "Spilled SGPRs: %d\n"
140 "Spilled VGPRs: %d\n"
141 "Code Size: %d bytes\n"
143 "Scratch: %d bytes per wave\n"
145 "********************\n\n\n",
146 conf
->num_sgprs
, conf
->num_vgprs
,
147 conf
->spilled_sgprs
, conf
->spilled_vgprs
, var
->code_size
,
148 conf
->lds_size
, conf
->scratch_bytes_per_wave
,
153 static struct radv_shader_variant
*
154 radv_pipeline_compile(struct radv_pipeline
*pipeline
,
155 struct radv_pipeline_cache
*cache
,
156 struct radv_shader_module
*module
,
157 const char *entrypoint
,
158 gl_shader_stage stage
,
159 const VkSpecializationInfo
*spec_info
,
160 struct radv_pipeline_layout
*layout
,
161 const struct ac_shader_variant_key
*key
)
163 unsigned char sha1
[20];
164 unsigned char gs_copy_sha1
[20];
165 struct radv_shader_variant
*variant
;
168 unsigned code_size
= 0;
171 _mesa_sha1_compute(module
->nir
->info
.name
,
172 strlen(module
->nir
->info
.name
),
175 radv_hash_shader(sha1
, module
, entrypoint
, spec_info
, layout
, key
, 0);
176 if (stage
== MESA_SHADER_GEOMETRY
)
177 radv_hash_shader(gs_copy_sha1
, module
, entrypoint
, spec_info
,
180 variant
= radv_create_shader_variant_from_pipeline_cache(pipeline
->device
,
184 if (stage
== MESA_SHADER_GEOMETRY
) {
185 pipeline
->gs_copy_shader
=
186 radv_create_shader_variant_from_pipeline_cache(
193 (stage
!= MESA_SHADER_GEOMETRY
|| pipeline
->gs_copy_shader
))
196 nir
= radv_shader_compile_to_nir(pipeline
->device
,
197 module
, entrypoint
, stage
,
203 variant
= radv_shader_variant_create(pipeline
->device
, nir
,
208 if (stage
== MESA_SHADER_GEOMETRY
&& !pipeline
->gs_copy_shader
) {
209 void *gs_copy_code
= NULL
;
210 unsigned gs_copy_code_size
= 0;
211 pipeline
->gs_copy_shader
= radv_create_gs_copy_shader(
212 pipeline
->device
, nir
, &gs_copy_code
,
213 &gs_copy_code_size
, key
->has_multiview_view_index
);
215 if (pipeline
->gs_copy_shader
) {
216 pipeline
->gs_copy_shader
=
217 radv_pipeline_cache_insert_shader(cache
,
219 pipeline
->gs_copy_shader
,
230 variant
= radv_pipeline_cache_insert_shader(cache
, sha1
, variant
,
238 static struct ac_shader_variant_key
239 radv_compute_tes_key(bool as_es
, bool export_prim_id
)
241 struct ac_shader_variant_key key
;
242 memset(&key
, 0, sizeof(key
));
243 key
.tes
.as_es
= as_es
;
244 /* export prim id only happens when no geom shader */
246 key
.tes
.export_prim_id
= export_prim_id
;
250 static struct ac_shader_variant_key
251 radv_compute_tcs_key(unsigned primitive_mode
, unsigned input_vertices
)
253 struct ac_shader_variant_key key
;
254 memset(&key
, 0, sizeof(key
));
255 key
.tcs
.primitive_mode
= primitive_mode
;
256 key
.tcs
.input_vertices
= input_vertices
;
261 radv_tess_pipeline_compile(struct radv_pipeline
*pipeline
,
262 struct radv_pipeline_cache
*cache
,
263 struct radv_shader_module
*tcs_module
,
264 struct radv_shader_module
*tes_module
,
265 const char *tcs_entrypoint
,
266 const char *tes_entrypoint
,
267 const VkSpecializationInfo
*tcs_spec_info
,
268 const VkSpecializationInfo
*tes_spec_info
,
269 struct radv_pipeline_layout
*layout
,
270 unsigned input_vertices
,
273 unsigned char tcs_sha1
[20], tes_sha1
[20];
274 struct radv_shader_variant
*tes_variant
= NULL
, *tcs_variant
= NULL
;
275 nir_shader
*tes_nir
, *tcs_nir
;
276 void *tes_code
= NULL
, *tcs_code
= NULL
;
277 unsigned tes_code_size
= 0, tcs_code_size
= 0;
278 struct ac_shader_variant_key tes_key
;
279 struct ac_shader_variant_key tcs_key
;
281 tes_key
= radv_compute_tes_key(radv_pipeline_has_gs(pipeline
),
282 pipeline
->shaders
[MESA_SHADER_FRAGMENT
]->info
.fs
.prim_id_input
);
283 tes_key
.has_multiview_view_index
= has_view_index
;
285 _mesa_sha1_compute(tes_module
->nir
->info
.name
,
286 strlen(tes_module
->nir
->info
.name
),
288 radv_hash_shader(tes_sha1
, tes_module
, tes_entrypoint
, tes_spec_info
, layout
, &tes_key
, 0);
290 tes_variant
= radv_create_shader_variant_from_pipeline_cache(pipeline
->device
,
295 tcs_key
= radv_compute_tcs_key(tes_variant
->info
.tes
.primitive_mode
, input_vertices
);
298 _mesa_sha1_compute(tcs_module
->nir
->info
.name
,
299 strlen(tcs_module
->nir
->info
.name
),
302 radv_hash_shader(tcs_sha1
, tcs_module
, tcs_entrypoint
, tcs_spec_info
, layout
, &tcs_key
, 0);
304 tcs_variant
= radv_create_shader_variant_from_pipeline_cache(pipeline
->device
,
309 if (tcs_variant
&& tes_variant
) {
310 pipeline
->shaders
[MESA_SHADER_TESS_CTRL
] = tcs_variant
;
311 pipeline
->shaders
[MESA_SHADER_TESS_EVAL
] = tes_variant
;
315 tes_nir
= radv_shader_compile_to_nir(pipeline
->device
,
316 tes_module
, tes_entrypoint
, MESA_SHADER_TESS_EVAL
,
321 tcs_nir
= radv_shader_compile_to_nir(pipeline
->device
,
322 tcs_module
, tcs_entrypoint
, MESA_SHADER_TESS_CTRL
,
327 nir_lower_tes_patch_vertices(tes_nir
,
328 tcs_nir
->info
.tess
.tcs_vertices_out
);
330 tes_variant
= radv_shader_variant_create(pipeline
->device
, tes_nir
,
331 layout
, &tes_key
, &tes_code
,
334 tcs_key
= radv_compute_tcs_key(tes_nir
->info
.tess
.primitive_mode
, input_vertices
);
336 _mesa_sha1_compute(tcs_module
->nir
->info
.name
,
337 strlen(tcs_module
->nir
->info
.name
),
340 radv_hash_shader(tcs_sha1
, tcs_module
, tcs_entrypoint
, tcs_spec_info
, layout
, &tcs_key
, 0);
342 tcs_variant
= radv_shader_variant_create(pipeline
->device
, tcs_nir
,
343 layout
, &tcs_key
, &tcs_code
,
346 if (!tes_module
->nir
)
347 ralloc_free(tes_nir
);
349 if (!tcs_module
->nir
)
350 ralloc_free(tcs_nir
);
353 tes_variant
= radv_pipeline_cache_insert_shader(cache
, tes_sha1
, tes_variant
,
354 tes_code
, tes_code_size
);
357 tcs_variant
= radv_pipeline_cache_insert_shader(cache
, tcs_sha1
, tcs_variant
,
358 tcs_code
, tcs_code_size
);
364 pipeline
->shaders
[MESA_SHADER_TESS_CTRL
] = tcs_variant
;
365 pipeline
->shaders
[MESA_SHADER_TESS_EVAL
] = tes_variant
;
370 radv_pipeline_scratch_init(struct radv_device
*device
,
371 struct radv_pipeline
*pipeline
)
373 unsigned scratch_bytes_per_wave
= 0;
374 unsigned max_waves
= 0;
375 unsigned min_waves
= 1;
377 for (int i
= 0; i
< MESA_SHADER_STAGES
; ++i
) {
378 if (pipeline
->shaders
[i
]) {
379 unsigned max_stage_waves
= device
->scratch_waves
;
381 scratch_bytes_per_wave
= MAX2(scratch_bytes_per_wave
,
382 pipeline
->shaders
[i
]->config
.scratch_bytes_per_wave
);
384 max_stage_waves
= MIN2(max_stage_waves
,
385 4 * device
->physical_device
->rad_info
.num_good_compute_units
*
386 (256 / pipeline
->shaders
[i
]->config
.num_vgprs
));
387 max_waves
= MAX2(max_waves
, max_stage_waves
);
391 if (pipeline
->shaders
[MESA_SHADER_COMPUTE
]) {
392 unsigned group_size
= pipeline
->shaders
[MESA_SHADER_COMPUTE
]->info
.cs
.block_size
[0] *
393 pipeline
->shaders
[MESA_SHADER_COMPUTE
]->info
.cs
.block_size
[1] *
394 pipeline
->shaders
[MESA_SHADER_COMPUTE
]->info
.cs
.block_size
[2];
395 min_waves
= MAX2(min_waves
, round_up_u32(group_size
, 64));
398 if (scratch_bytes_per_wave
)
399 max_waves
= MIN2(max_waves
, 0xffffffffu
/ scratch_bytes_per_wave
);
401 if (scratch_bytes_per_wave
&& max_waves
< min_waves
) {
402 /* Not really true at this moment, but will be true on first
403 * execution. Avoid having hanging shaders. */
404 return VK_ERROR_OUT_OF_DEVICE_MEMORY
;
406 pipeline
->scratch_bytes_per_wave
= scratch_bytes_per_wave
;
407 pipeline
->max_waves
= max_waves
;
411 static uint32_t si_translate_blend_function(VkBlendOp op
)
414 case VK_BLEND_OP_ADD
:
415 return V_028780_COMB_DST_PLUS_SRC
;
416 case VK_BLEND_OP_SUBTRACT
:
417 return V_028780_COMB_SRC_MINUS_DST
;
418 case VK_BLEND_OP_REVERSE_SUBTRACT
:
419 return V_028780_COMB_DST_MINUS_SRC
;
420 case VK_BLEND_OP_MIN
:
421 return V_028780_COMB_MIN_DST_SRC
;
422 case VK_BLEND_OP_MAX
:
423 return V_028780_COMB_MAX_DST_SRC
;
429 static uint32_t si_translate_blend_factor(VkBlendFactor factor
)
432 case VK_BLEND_FACTOR_ZERO
:
433 return V_028780_BLEND_ZERO
;
434 case VK_BLEND_FACTOR_ONE
:
435 return V_028780_BLEND_ONE
;
436 case VK_BLEND_FACTOR_SRC_COLOR
:
437 return V_028780_BLEND_SRC_COLOR
;
438 case VK_BLEND_FACTOR_ONE_MINUS_SRC_COLOR
:
439 return V_028780_BLEND_ONE_MINUS_SRC_COLOR
;
440 case VK_BLEND_FACTOR_DST_COLOR
:
441 return V_028780_BLEND_DST_COLOR
;
442 case VK_BLEND_FACTOR_ONE_MINUS_DST_COLOR
:
443 return V_028780_BLEND_ONE_MINUS_DST_COLOR
;
444 case VK_BLEND_FACTOR_SRC_ALPHA
:
445 return V_028780_BLEND_SRC_ALPHA
;
446 case VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA
:
447 return V_028780_BLEND_ONE_MINUS_SRC_ALPHA
;
448 case VK_BLEND_FACTOR_DST_ALPHA
:
449 return V_028780_BLEND_DST_ALPHA
;
450 case VK_BLEND_FACTOR_ONE_MINUS_DST_ALPHA
:
451 return V_028780_BLEND_ONE_MINUS_DST_ALPHA
;
452 case VK_BLEND_FACTOR_CONSTANT_COLOR
:
453 return V_028780_BLEND_CONSTANT_COLOR
;
454 case VK_BLEND_FACTOR_ONE_MINUS_CONSTANT_COLOR
:
455 return V_028780_BLEND_ONE_MINUS_CONSTANT_COLOR
;
456 case VK_BLEND_FACTOR_CONSTANT_ALPHA
:
457 return V_028780_BLEND_CONSTANT_ALPHA
;
458 case VK_BLEND_FACTOR_ONE_MINUS_CONSTANT_ALPHA
:
459 return V_028780_BLEND_ONE_MINUS_CONSTANT_ALPHA
;
460 case VK_BLEND_FACTOR_SRC_ALPHA_SATURATE
:
461 return V_028780_BLEND_SRC_ALPHA_SATURATE
;
462 case VK_BLEND_FACTOR_SRC1_COLOR
:
463 return V_028780_BLEND_SRC1_COLOR
;
464 case VK_BLEND_FACTOR_ONE_MINUS_SRC1_COLOR
:
465 return V_028780_BLEND_INV_SRC1_COLOR
;
466 case VK_BLEND_FACTOR_SRC1_ALPHA
:
467 return V_028780_BLEND_SRC1_ALPHA
;
468 case VK_BLEND_FACTOR_ONE_MINUS_SRC1_ALPHA
:
469 return V_028780_BLEND_INV_SRC1_ALPHA
;
475 static uint32_t si_translate_blend_opt_function(VkBlendOp op
)
478 case VK_BLEND_OP_ADD
:
479 return V_028760_OPT_COMB_ADD
;
480 case VK_BLEND_OP_SUBTRACT
:
481 return V_028760_OPT_COMB_SUBTRACT
;
482 case VK_BLEND_OP_REVERSE_SUBTRACT
:
483 return V_028760_OPT_COMB_REVSUBTRACT
;
484 case VK_BLEND_OP_MIN
:
485 return V_028760_OPT_COMB_MIN
;
486 case VK_BLEND_OP_MAX
:
487 return V_028760_OPT_COMB_MAX
;
489 return V_028760_OPT_COMB_BLEND_DISABLED
;
493 static uint32_t si_translate_blend_opt_factor(VkBlendFactor factor
, bool is_alpha
)
496 case VK_BLEND_FACTOR_ZERO
:
497 return V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_ALL
;
498 case VK_BLEND_FACTOR_ONE
:
499 return V_028760_BLEND_OPT_PRESERVE_ALL_IGNORE_NONE
;
500 case VK_BLEND_FACTOR_SRC_COLOR
:
501 return is_alpha
? V_028760_BLEND_OPT_PRESERVE_A1_IGNORE_A0
502 : V_028760_BLEND_OPT_PRESERVE_C1_IGNORE_C0
;
503 case VK_BLEND_FACTOR_ONE_MINUS_SRC_COLOR
:
504 return is_alpha
? V_028760_BLEND_OPT_PRESERVE_A0_IGNORE_A1
505 : V_028760_BLEND_OPT_PRESERVE_C0_IGNORE_C1
;
506 case VK_BLEND_FACTOR_SRC_ALPHA
:
507 return V_028760_BLEND_OPT_PRESERVE_A1_IGNORE_A0
;
508 case VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA
:
509 return V_028760_BLEND_OPT_PRESERVE_A0_IGNORE_A1
;
510 case VK_BLEND_FACTOR_SRC_ALPHA_SATURATE
:
511 return is_alpha
? V_028760_BLEND_OPT_PRESERVE_ALL_IGNORE_NONE
512 : V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_A0
;
514 return V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE
;
519 * Get rid of DST in the blend factors by commuting the operands:
520 * func(src * DST, dst * 0) ---> func(src * 0, dst * SRC)
522 static void si_blend_remove_dst(unsigned *func
, unsigned *src_factor
,
523 unsigned *dst_factor
, unsigned expected_dst
,
524 unsigned replacement_src
)
526 if (*src_factor
== expected_dst
&&
527 *dst_factor
== VK_BLEND_FACTOR_ZERO
) {
528 *src_factor
= VK_BLEND_FACTOR_ZERO
;
529 *dst_factor
= replacement_src
;
531 /* Commuting the operands requires reversing subtractions. */
532 if (*func
== VK_BLEND_OP_SUBTRACT
)
533 *func
= VK_BLEND_OP_REVERSE_SUBTRACT
;
534 else if (*func
== VK_BLEND_OP_REVERSE_SUBTRACT
)
535 *func
= VK_BLEND_OP_SUBTRACT
;
539 static bool si_blend_factor_uses_dst(unsigned factor
)
541 return factor
== VK_BLEND_FACTOR_DST_COLOR
||
542 factor
== VK_BLEND_FACTOR_DST_ALPHA
||
543 factor
== VK_BLEND_FACTOR_SRC_ALPHA_SATURATE
||
544 factor
== VK_BLEND_FACTOR_ONE_MINUS_DST_ALPHA
||
545 factor
== VK_BLEND_FACTOR_ONE_MINUS_DST_COLOR
;
548 static bool is_dual_src(VkBlendFactor factor
)
551 case VK_BLEND_FACTOR_SRC1_COLOR
:
552 case VK_BLEND_FACTOR_ONE_MINUS_SRC1_COLOR
:
553 case VK_BLEND_FACTOR_SRC1_ALPHA
:
554 case VK_BLEND_FACTOR_ONE_MINUS_SRC1_ALPHA
:
561 static unsigned si_choose_spi_color_format(VkFormat vk_format
,
563 bool blend_need_alpha
)
565 const struct vk_format_description
*desc
= vk_format_description(vk_format
);
566 unsigned format
, ntype
, swap
;
568 /* Alpha is needed for alpha-to-coverage.
569 * Blending may be with or without alpha.
571 unsigned normal
= 0; /* most optimal, may not support blending or export alpha */
572 unsigned alpha
= 0; /* exports alpha, but may not support blending */
573 unsigned blend
= 0; /* supports blending, but may not export alpha */
574 unsigned blend_alpha
= 0; /* least optimal, supports blending and exports alpha */
576 format
= radv_translate_colorformat(vk_format
);
577 ntype
= radv_translate_color_numformat(vk_format
, desc
,
578 vk_format_get_first_non_void_channel(vk_format
));
579 swap
= radv_translate_colorswap(vk_format
, false);
581 /* Choose the SPI color formats. These are required values for Stoney/RB+.
582 * Other chips have multiple choices, though they are not necessarily better.
585 case V_028C70_COLOR_5_6_5
:
586 case V_028C70_COLOR_1_5_5_5
:
587 case V_028C70_COLOR_5_5_5_1
:
588 case V_028C70_COLOR_4_4_4_4
:
589 case V_028C70_COLOR_10_11_11
:
590 case V_028C70_COLOR_11_11_10
:
591 case V_028C70_COLOR_8
:
592 case V_028C70_COLOR_8_8
:
593 case V_028C70_COLOR_8_8_8_8
:
594 case V_028C70_COLOR_10_10_10_2
:
595 case V_028C70_COLOR_2_10_10_10
:
596 if (ntype
== V_028C70_NUMBER_UINT
)
597 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_UINT16_ABGR
;
598 else if (ntype
== V_028C70_NUMBER_SINT
)
599 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_SINT16_ABGR
;
601 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_FP16_ABGR
;
604 case V_028C70_COLOR_16
:
605 case V_028C70_COLOR_16_16
:
606 case V_028C70_COLOR_16_16_16_16
:
607 if (ntype
== V_028C70_NUMBER_UNORM
||
608 ntype
== V_028C70_NUMBER_SNORM
) {
609 /* UNORM16 and SNORM16 don't support blending */
610 if (ntype
== V_028C70_NUMBER_UNORM
)
611 normal
= alpha
= V_028714_SPI_SHADER_UNORM16_ABGR
;
613 normal
= alpha
= V_028714_SPI_SHADER_SNORM16_ABGR
;
615 /* Use 32 bits per channel for blending. */
616 if (format
== V_028C70_COLOR_16
) {
617 if (swap
== V_028C70_SWAP_STD
) { /* R */
618 blend
= V_028714_SPI_SHADER_32_R
;
619 blend_alpha
= V_028714_SPI_SHADER_32_AR
;
620 } else if (swap
== V_028C70_SWAP_ALT_REV
) /* A */
621 blend
= blend_alpha
= V_028714_SPI_SHADER_32_AR
;
624 } else if (format
== V_028C70_COLOR_16_16
) {
625 if (swap
== V_028C70_SWAP_STD
) { /* RG */
626 blend
= V_028714_SPI_SHADER_32_GR
;
627 blend_alpha
= V_028714_SPI_SHADER_32_ABGR
;
628 } else if (swap
== V_028C70_SWAP_ALT
) /* RA */
629 blend
= blend_alpha
= V_028714_SPI_SHADER_32_AR
;
632 } else /* 16_16_16_16 */
633 blend
= blend_alpha
= V_028714_SPI_SHADER_32_ABGR
;
634 } else if (ntype
== V_028C70_NUMBER_UINT
)
635 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_UINT16_ABGR
;
636 else if (ntype
== V_028C70_NUMBER_SINT
)
637 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_SINT16_ABGR
;
638 else if (ntype
== V_028C70_NUMBER_FLOAT
)
639 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_FP16_ABGR
;
644 case V_028C70_COLOR_32
:
645 if (swap
== V_028C70_SWAP_STD
) { /* R */
646 blend
= normal
= V_028714_SPI_SHADER_32_R
;
647 alpha
= blend_alpha
= V_028714_SPI_SHADER_32_AR
;
648 } else if (swap
== V_028C70_SWAP_ALT_REV
) /* A */
649 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_32_AR
;
654 case V_028C70_COLOR_32_32
:
655 if (swap
== V_028C70_SWAP_STD
) { /* RG */
656 blend
= normal
= V_028714_SPI_SHADER_32_GR
;
657 alpha
= blend_alpha
= V_028714_SPI_SHADER_32_ABGR
;
658 } else if (swap
== V_028C70_SWAP_ALT
) /* RA */
659 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_32_AR
;
664 case V_028C70_COLOR_32_32_32_32
:
665 case V_028C70_COLOR_8_24
:
666 case V_028C70_COLOR_24_8
:
667 case V_028C70_COLOR_X24_8_32_FLOAT
:
668 alpha
= blend
= blend_alpha
= normal
= V_028714_SPI_SHADER_32_ABGR
;
672 unreachable("unhandled blend format");
675 if (blend_enable
&& blend_need_alpha
)
677 else if(blend_need_alpha
)
679 else if(blend_enable
)
685 static unsigned si_get_cb_shader_mask(unsigned spi_shader_col_format
)
687 unsigned i
, cb_shader_mask
= 0;
689 for (i
= 0; i
< 8; i
++) {
690 switch ((spi_shader_col_format
>> (i
* 4)) & 0xf) {
691 case V_028714_SPI_SHADER_ZERO
:
693 case V_028714_SPI_SHADER_32_R
:
694 cb_shader_mask
|= 0x1 << (i
* 4);
696 case V_028714_SPI_SHADER_32_GR
:
697 cb_shader_mask
|= 0x3 << (i
* 4);
699 case V_028714_SPI_SHADER_32_AR
:
700 cb_shader_mask
|= 0x9 << (i
* 4);
702 case V_028714_SPI_SHADER_FP16_ABGR
:
703 case V_028714_SPI_SHADER_UNORM16_ABGR
:
704 case V_028714_SPI_SHADER_SNORM16_ABGR
:
705 case V_028714_SPI_SHADER_UINT16_ABGR
:
706 case V_028714_SPI_SHADER_SINT16_ABGR
:
707 case V_028714_SPI_SHADER_32_ABGR
:
708 cb_shader_mask
|= 0xf << (i
* 4);
714 return cb_shader_mask
;
718 radv_pipeline_compute_spi_color_formats(struct radv_pipeline
*pipeline
,
719 const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
720 uint32_t blend_enable
,
721 uint32_t blend_need_alpha
,
722 bool single_cb_enable
,
723 bool blend_mrt0_is_dual_src
)
725 RADV_FROM_HANDLE(radv_render_pass
, pass
, pCreateInfo
->renderPass
);
726 struct radv_subpass
*subpass
= pass
->subpasses
+ pCreateInfo
->subpass
;
727 struct radv_blend_state
*blend
= &pipeline
->graphics
.blend
;
728 unsigned col_format
= 0;
730 for (unsigned i
= 0; i
< (single_cb_enable
? 1 : subpass
->color_count
); ++i
) {
733 if (subpass
->color_attachments
[i
].attachment
== VK_ATTACHMENT_UNUSED
) {
734 cf
= V_028714_SPI_SHADER_ZERO
;
736 struct radv_render_pass_attachment
*attachment
= pass
->attachments
+ subpass
->color_attachments
[i
].attachment
;
738 cf
= si_choose_spi_color_format(attachment
->format
,
739 blend_enable
& (1 << i
),
740 blend_need_alpha
& (1 << i
));
743 col_format
|= cf
<< (4 * i
);
746 blend
->cb_shader_mask
= si_get_cb_shader_mask(col_format
);
748 if (blend_mrt0_is_dual_src
)
749 col_format
|= (col_format
& 0xf) << 4;
750 blend
->spi_shader_col_format
= col_format
;
754 format_is_int8(VkFormat format
)
756 const struct vk_format_description
*desc
= vk_format_description(format
);
757 int channel
= vk_format_get_first_non_void_channel(format
);
759 return channel
>= 0 && desc
->channel
[channel
].pure_integer
&&
760 desc
->channel
[channel
].size
== 8;
764 format_is_int10(VkFormat format
)
766 const struct vk_format_description
*desc
= vk_format_description(format
);
768 if (desc
->nr_channels
!= 4)
770 for (unsigned i
= 0; i
< 4; i
++) {
771 if (desc
->channel
[i
].pure_integer
&& desc
->channel
[i
].size
== 10)
777 unsigned radv_format_meta_fs_key(VkFormat format
)
779 unsigned col_format
= si_choose_spi_color_format(format
, false, false) - 1;
780 bool is_int8
= format_is_int8(format
);
781 bool is_int10
= format_is_int10(format
);
783 return col_format
+ (is_int8
? 3 : is_int10
? 5 : 0);
787 radv_pipeline_compute_get_int_clamp(const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
788 unsigned *is_int8
, unsigned *is_int10
)
790 RADV_FROM_HANDLE(radv_render_pass
, pass
, pCreateInfo
->renderPass
);
791 struct radv_subpass
*subpass
= pass
->subpasses
+ pCreateInfo
->subpass
;
795 for (unsigned i
= 0; i
< subpass
->color_count
; ++i
) {
796 struct radv_render_pass_attachment
*attachment
;
798 if (subpass
->color_attachments
[i
].attachment
== VK_ATTACHMENT_UNUSED
)
801 attachment
= pass
->attachments
+ subpass
->color_attachments
[i
].attachment
;
803 if (format_is_int8(attachment
->format
))
805 if (format_is_int10(attachment
->format
))
811 radv_pipeline_init_blend_state(struct radv_pipeline
*pipeline
,
812 const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
813 const struct radv_graphics_pipeline_create_info
*extra
)
815 const VkPipelineColorBlendStateCreateInfo
*vkblend
= pCreateInfo
->pColorBlendState
;
816 struct radv_blend_state
*blend
= &pipeline
->graphics
.blend
;
817 unsigned mode
= V_028808_CB_NORMAL
;
818 uint32_t blend_enable
= 0, blend_need_alpha
= 0;
819 bool blend_mrt0_is_dual_src
= false;
821 bool single_cb_enable
= false;
826 if (extra
&& extra
->custom_blend_mode
) {
827 single_cb_enable
= true;
828 mode
= extra
->custom_blend_mode
;
830 blend
->cb_color_control
= 0;
831 if (vkblend
->logicOpEnable
)
832 blend
->cb_color_control
|= S_028808_ROP3(vkblend
->logicOp
| (vkblend
->logicOp
<< 4));
834 blend
->cb_color_control
|= S_028808_ROP3(0xcc);
836 blend
->db_alpha_to_mask
= S_028B70_ALPHA_TO_MASK_OFFSET0(2) |
837 S_028B70_ALPHA_TO_MASK_OFFSET1(2) |
838 S_028B70_ALPHA_TO_MASK_OFFSET2(2) |
839 S_028B70_ALPHA_TO_MASK_OFFSET3(2);
841 blend
->cb_target_mask
= 0;
842 for (i
= 0; i
< vkblend
->attachmentCount
; i
++) {
843 const VkPipelineColorBlendAttachmentState
*att
= &vkblend
->pAttachments
[i
];
844 unsigned blend_cntl
= 0;
845 unsigned srcRGB_opt
, dstRGB_opt
, srcA_opt
, dstA_opt
;
846 VkBlendOp eqRGB
= att
->colorBlendOp
;
847 VkBlendFactor srcRGB
= att
->srcColorBlendFactor
;
848 VkBlendFactor dstRGB
= att
->dstColorBlendFactor
;
849 VkBlendOp eqA
= att
->alphaBlendOp
;
850 VkBlendFactor srcA
= att
->srcAlphaBlendFactor
;
851 VkBlendFactor dstA
= att
->dstAlphaBlendFactor
;
853 blend
->sx_mrt_blend_opt
[i
] = S_028760_COLOR_COMB_FCN(V_028760_OPT_COMB_BLEND_DISABLED
) | S_028760_ALPHA_COMB_FCN(V_028760_OPT_COMB_BLEND_DISABLED
);
855 if (!att
->colorWriteMask
)
858 blend
->cb_target_mask
|= (unsigned)att
->colorWriteMask
<< (4 * i
);
859 if (!att
->blendEnable
) {
860 blend
->cb_blend_control
[i
] = blend_cntl
;
864 if (is_dual_src(srcRGB
) || is_dual_src(dstRGB
) || is_dual_src(srcA
) || is_dual_src(dstA
))
866 blend_mrt0_is_dual_src
= true;
868 if (eqRGB
== VK_BLEND_OP_MIN
|| eqRGB
== VK_BLEND_OP_MAX
) {
869 srcRGB
= VK_BLEND_FACTOR_ONE
;
870 dstRGB
= VK_BLEND_FACTOR_ONE
;
872 if (eqA
== VK_BLEND_OP_MIN
|| eqA
== VK_BLEND_OP_MAX
) {
873 srcA
= VK_BLEND_FACTOR_ONE
;
874 dstA
= VK_BLEND_FACTOR_ONE
;
877 /* Blending optimizations for RB+.
878 * These transformations don't change the behavior.
880 * First, get rid of DST in the blend factors:
881 * func(src * DST, dst * 0) ---> func(src * 0, dst * SRC)
883 si_blend_remove_dst(&eqRGB
, &srcRGB
, &dstRGB
,
884 VK_BLEND_FACTOR_DST_COLOR
,
885 VK_BLEND_FACTOR_SRC_COLOR
);
887 si_blend_remove_dst(&eqA
, &srcA
, &dstA
,
888 VK_BLEND_FACTOR_DST_COLOR
,
889 VK_BLEND_FACTOR_SRC_COLOR
);
891 si_blend_remove_dst(&eqA
, &srcA
, &dstA
,
892 VK_BLEND_FACTOR_DST_ALPHA
,
893 VK_BLEND_FACTOR_SRC_ALPHA
);
895 /* Look up the ideal settings from tables. */
896 srcRGB_opt
= si_translate_blend_opt_factor(srcRGB
, false);
897 dstRGB_opt
= si_translate_blend_opt_factor(dstRGB
, false);
898 srcA_opt
= si_translate_blend_opt_factor(srcA
, true);
899 dstA_opt
= si_translate_blend_opt_factor(dstA
, true);
901 /* Handle interdependencies. */
902 if (si_blend_factor_uses_dst(srcRGB
))
903 dstRGB_opt
= V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE
;
904 if (si_blend_factor_uses_dst(srcA
))
905 dstA_opt
= V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_NONE
;
907 if (srcRGB
== VK_BLEND_FACTOR_SRC_ALPHA_SATURATE
&&
908 (dstRGB
== VK_BLEND_FACTOR_ZERO
||
909 dstRGB
== VK_BLEND_FACTOR_SRC_ALPHA
||
910 dstRGB
== VK_BLEND_FACTOR_SRC_ALPHA_SATURATE
))
911 dstRGB_opt
= V_028760_BLEND_OPT_PRESERVE_NONE_IGNORE_A0
;
913 /* Set the final value. */
914 blend
->sx_mrt_blend_opt
[i
] =
915 S_028760_COLOR_SRC_OPT(srcRGB_opt
) |
916 S_028760_COLOR_DST_OPT(dstRGB_opt
) |
917 S_028760_COLOR_COMB_FCN(si_translate_blend_opt_function(eqRGB
)) |
918 S_028760_ALPHA_SRC_OPT(srcA_opt
) |
919 S_028760_ALPHA_DST_OPT(dstA_opt
) |
920 S_028760_ALPHA_COMB_FCN(si_translate_blend_opt_function(eqA
));
921 blend_cntl
|= S_028780_ENABLE(1);
923 blend_cntl
|= S_028780_COLOR_COMB_FCN(si_translate_blend_function(eqRGB
));
924 blend_cntl
|= S_028780_COLOR_SRCBLEND(si_translate_blend_factor(srcRGB
));
925 blend_cntl
|= S_028780_COLOR_DESTBLEND(si_translate_blend_factor(dstRGB
));
926 if (srcA
!= srcRGB
|| dstA
!= dstRGB
|| eqA
!= eqRGB
) {
927 blend_cntl
|= S_028780_SEPARATE_ALPHA_BLEND(1);
928 blend_cntl
|= S_028780_ALPHA_COMB_FCN(si_translate_blend_function(eqA
));
929 blend_cntl
|= S_028780_ALPHA_SRCBLEND(si_translate_blend_factor(srcA
));
930 blend_cntl
|= S_028780_ALPHA_DESTBLEND(si_translate_blend_factor(dstA
));
932 blend
->cb_blend_control
[i
] = blend_cntl
;
934 blend_enable
|= 1 << i
;
936 if (srcRGB
== VK_BLEND_FACTOR_SRC_ALPHA
||
937 dstRGB
== VK_BLEND_FACTOR_SRC_ALPHA
||
938 srcRGB
== VK_BLEND_FACTOR_SRC_ALPHA_SATURATE
||
939 dstRGB
== VK_BLEND_FACTOR_SRC_ALPHA_SATURATE
||
940 srcRGB
== VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA
||
941 dstRGB
== VK_BLEND_FACTOR_ONE_MINUS_SRC_ALPHA
)
942 blend_need_alpha
|= 1 << i
;
944 for (i
= vkblend
->attachmentCount
; i
< 8; i
++) {
945 blend
->cb_blend_control
[i
] = 0;
946 blend
->sx_mrt_blend_opt
[i
] = S_028760_COLOR_COMB_FCN(V_028760_OPT_COMB_BLEND_DISABLED
) | S_028760_ALPHA_COMB_FCN(V_028760_OPT_COMB_BLEND_DISABLED
);
949 /* disable RB+ for now */
950 if (pipeline
->device
->physical_device
->has_rbplus
)
951 blend
->cb_color_control
|= S_028808_DISABLE_DUAL_QUAD(1);
953 if (blend
->cb_target_mask
)
954 blend
->cb_color_control
|= S_028808_MODE(mode
);
956 blend
->cb_color_control
|= S_028808_MODE(V_028808_CB_DISABLE
);
958 radv_pipeline_compute_spi_color_formats(pipeline
, pCreateInfo
,
959 blend_enable
, blend_need_alpha
, single_cb_enable
, blend_mrt0_is_dual_src
);
962 static uint32_t si_translate_stencil_op(enum VkStencilOp op
)
965 case VK_STENCIL_OP_KEEP
:
966 return V_02842C_STENCIL_KEEP
;
967 case VK_STENCIL_OP_ZERO
:
968 return V_02842C_STENCIL_ZERO
;
969 case VK_STENCIL_OP_REPLACE
:
970 return V_02842C_STENCIL_REPLACE_TEST
;
971 case VK_STENCIL_OP_INCREMENT_AND_CLAMP
:
972 return V_02842C_STENCIL_ADD_CLAMP
;
973 case VK_STENCIL_OP_DECREMENT_AND_CLAMP
:
974 return V_02842C_STENCIL_SUB_CLAMP
;
975 case VK_STENCIL_OP_INVERT
:
976 return V_02842C_STENCIL_INVERT
;
977 case VK_STENCIL_OP_INCREMENT_AND_WRAP
:
978 return V_02842C_STENCIL_ADD_WRAP
;
979 case VK_STENCIL_OP_DECREMENT_AND_WRAP
:
980 return V_02842C_STENCIL_SUB_WRAP
;
986 radv_pipeline_init_depth_stencil_state(struct radv_pipeline
*pipeline
,
987 const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
988 const struct radv_graphics_pipeline_create_info
*extra
)
990 const VkPipelineDepthStencilStateCreateInfo
*vkds
= pCreateInfo
->pDepthStencilState
;
991 struct radv_depth_stencil_state
*ds
= &pipeline
->graphics
.ds
;
993 memset(ds
, 0, sizeof(*ds
));
997 RADV_FROM_HANDLE(radv_render_pass
, pass
, pCreateInfo
->renderPass
);
998 struct radv_subpass
*subpass
= pass
->subpasses
+ pCreateInfo
->subpass
;
999 if (subpass
->depth_stencil_attachment
.attachment
== VK_ATTACHMENT_UNUSED
)
1002 struct radv_render_pass_attachment
*attachment
= pass
->attachments
+ subpass
->depth_stencil_attachment
.attachment
;
1003 bool has_depth_attachment
= vk_format_is_depth(attachment
->format
);
1004 bool has_stencil_attachment
= vk_format_is_stencil(attachment
->format
);
1006 if (has_depth_attachment
) {
1007 ds
->db_depth_control
= S_028800_Z_ENABLE(vkds
->depthTestEnable
? 1 : 0) |
1008 S_028800_Z_WRITE_ENABLE(vkds
->depthWriteEnable
? 1 : 0) |
1009 S_028800_ZFUNC(vkds
->depthCompareOp
) |
1010 S_028800_DEPTH_BOUNDS_ENABLE(vkds
->depthBoundsTestEnable
? 1 : 0);
1013 if (has_stencil_attachment
&& vkds
->stencilTestEnable
) {
1014 ds
->db_depth_control
|= S_028800_STENCIL_ENABLE(1) | S_028800_BACKFACE_ENABLE(1);
1015 ds
->db_depth_control
|= S_028800_STENCILFUNC(vkds
->front
.compareOp
);
1016 ds
->db_stencil_control
|= S_02842C_STENCILFAIL(si_translate_stencil_op(vkds
->front
.failOp
));
1017 ds
->db_stencil_control
|= S_02842C_STENCILZPASS(si_translate_stencil_op(vkds
->front
.passOp
));
1018 ds
->db_stencil_control
|= S_02842C_STENCILZFAIL(si_translate_stencil_op(vkds
->front
.depthFailOp
));
1020 ds
->db_depth_control
|= S_028800_STENCILFUNC_BF(vkds
->back
.compareOp
);
1021 ds
->db_stencil_control
|= S_02842C_STENCILFAIL_BF(si_translate_stencil_op(vkds
->back
.failOp
));
1022 ds
->db_stencil_control
|= S_02842C_STENCILZPASS_BF(si_translate_stencil_op(vkds
->back
.passOp
));
1023 ds
->db_stencil_control
|= S_02842C_STENCILZFAIL_BF(si_translate_stencil_op(vkds
->back
.depthFailOp
));
1028 ds
->db_render_control
|= S_028000_DEPTH_CLEAR_ENABLE(extra
->db_depth_clear
);
1029 ds
->db_render_control
|= S_028000_STENCIL_CLEAR_ENABLE(extra
->db_stencil_clear
);
1031 ds
->db_render_control
|= S_028000_RESUMMARIZE_ENABLE(extra
->db_resummarize
);
1032 ds
->db_render_control
|= S_028000_DEPTH_COMPRESS_DISABLE(extra
->db_flush_depth_inplace
);
1033 ds
->db_render_control
|= S_028000_STENCIL_COMPRESS_DISABLE(extra
->db_flush_stencil_inplace
);
1034 ds
->db_render_override2
|= S_028010_DISABLE_ZMASK_EXPCLEAR_OPTIMIZATION(extra
->db_depth_disable_expclear
);
1035 ds
->db_render_override2
|= S_028010_DISABLE_SMEM_EXPCLEAR_OPTIMIZATION(extra
->db_stencil_disable_expclear
);
1039 static uint32_t si_translate_fill(VkPolygonMode func
)
1042 case VK_POLYGON_MODE_FILL
:
1043 return V_028814_X_DRAW_TRIANGLES
;
1044 case VK_POLYGON_MODE_LINE
:
1045 return V_028814_X_DRAW_LINES
;
1046 case VK_POLYGON_MODE_POINT
:
1047 return V_028814_X_DRAW_POINTS
;
1050 return V_028814_X_DRAW_POINTS
;
1054 radv_pipeline_init_raster_state(struct radv_pipeline
*pipeline
,
1055 const VkGraphicsPipelineCreateInfo
*pCreateInfo
)
1057 const VkPipelineRasterizationStateCreateInfo
*vkraster
= pCreateInfo
->pRasterizationState
;
1058 struct radv_raster_state
*raster
= &pipeline
->graphics
.raster
;
1060 memset(raster
, 0, sizeof(*raster
));
1062 raster
->spi_interp_control
=
1063 S_0286D4_FLAT_SHADE_ENA(1) |
1064 S_0286D4_PNT_SPRITE_ENA(1) |
1065 S_0286D4_PNT_SPRITE_OVRD_X(V_0286D4_SPI_PNT_SPRITE_SEL_S
) |
1066 S_0286D4_PNT_SPRITE_OVRD_Y(V_0286D4_SPI_PNT_SPRITE_SEL_T
) |
1067 S_0286D4_PNT_SPRITE_OVRD_Z(V_0286D4_SPI_PNT_SPRITE_SEL_0
) |
1068 S_0286D4_PNT_SPRITE_OVRD_W(V_0286D4_SPI_PNT_SPRITE_SEL_1
) |
1069 S_0286D4_PNT_SPRITE_TOP_1(0); // vulkan is top to bottom - 1.0 at bottom
1072 raster
->pa_cl_clip_cntl
= S_028810_PS_UCP_MODE(3) |
1073 S_028810_DX_CLIP_SPACE_DEF(1) | // vulkan uses DX conventions.
1074 S_028810_ZCLIP_NEAR_DISABLE(vkraster
->depthClampEnable
? 1 : 0) |
1075 S_028810_ZCLIP_FAR_DISABLE(vkraster
->depthClampEnable
? 1 : 0) |
1076 S_028810_DX_RASTERIZATION_KILL(vkraster
->rasterizerDiscardEnable
? 1 : 0) |
1077 S_028810_DX_LINEAR_ATTR_CLIP_ENA(1);
1079 raster
->pa_su_vtx_cntl
=
1080 S_028BE4_PIX_CENTER(1) | // TODO verify
1081 S_028BE4_ROUND_MODE(V_028BE4_X_ROUND_TO_EVEN
) |
1082 S_028BE4_QUANT_MODE(V_028BE4_X_16_8_FIXED_POINT_1_256TH
);
1084 raster
->pa_su_sc_mode_cntl
=
1085 S_028814_FACE(vkraster
->frontFace
) |
1086 S_028814_CULL_FRONT(!!(vkraster
->cullMode
& VK_CULL_MODE_FRONT_BIT
)) |
1087 S_028814_CULL_BACK(!!(vkraster
->cullMode
& VK_CULL_MODE_BACK_BIT
)) |
1088 S_028814_POLY_MODE(vkraster
->polygonMode
!= VK_POLYGON_MODE_FILL
) |
1089 S_028814_POLYMODE_FRONT_PTYPE(si_translate_fill(vkraster
->polygonMode
)) |
1090 S_028814_POLYMODE_BACK_PTYPE(si_translate_fill(vkraster
->polygonMode
)) |
1091 S_028814_POLY_OFFSET_FRONT_ENABLE(vkraster
->depthBiasEnable
? 1 : 0) |
1092 S_028814_POLY_OFFSET_BACK_ENABLE(vkraster
->depthBiasEnable
? 1 : 0) |
1093 S_028814_POLY_OFFSET_PARA_ENABLE(vkraster
->depthBiasEnable
? 1 : 0);
1098 radv_pipeline_init_multisample_state(struct radv_pipeline
*pipeline
,
1099 const VkGraphicsPipelineCreateInfo
*pCreateInfo
)
1101 const VkPipelineMultisampleStateCreateInfo
*vkms
= pCreateInfo
->pMultisampleState
;
1102 struct radv_blend_state
*blend
= &pipeline
->graphics
.blend
;
1103 struct radv_multisample_state
*ms
= &pipeline
->graphics
.ms
;
1104 unsigned num_tile_pipes
= pipeline
->device
->physical_device
->rad_info
.num_tile_pipes
;
1105 int ps_iter_samples
= 1;
1106 uint32_t mask
= 0xffff;
1109 ms
->num_samples
= vkms
->rasterizationSamples
;
1111 ms
->num_samples
= 1;
1113 if (vkms
&& vkms
->sampleShadingEnable
) {
1114 ps_iter_samples
= ceil(vkms
->minSampleShading
* ms
->num_samples
);
1115 } else if (pipeline
->shaders
[MESA_SHADER_FRAGMENT
]->info
.info
.ps
.force_persample
) {
1116 ps_iter_samples
= ms
->num_samples
;
1119 ms
->pa_sc_line_cntl
= S_028BDC_DX10_DIAMOND_TEST_ENA(1);
1120 ms
->pa_sc_aa_config
= 0;
1121 ms
->db_eqaa
= S_028804_HIGH_QUALITY_INTERSECTIONS(1) |
1122 S_028804_STATIC_ANCHOR_ASSOCIATIONS(1);
1123 ms
->pa_sc_mode_cntl_1
=
1124 S_028A4C_WALK_FENCE_ENABLE(1) | //TODO linear dst fixes
1125 S_028A4C_WALK_FENCE_SIZE(num_tile_pipes
== 2 ? 2 : 3) |
1127 S_028A4C_WALK_ALIGN8_PRIM_FITS_ST(1) |
1128 S_028A4C_SUPERTILE_WALK_ORDER_ENABLE(1) |
1129 S_028A4C_TILE_WALK_ORDER_ENABLE(1) |
1130 S_028A4C_MULTI_SHADER_ENGINE_PRIM_DISCARD_ENABLE(1) |
1131 EG_S_028A4C_FORCE_EOV_CNTDWN_ENABLE(1) |
1132 EG_S_028A4C_FORCE_EOV_REZ_ENABLE(1);
1133 ms
->pa_sc_mode_cntl_0
= S_028A48_ALTERNATE_RBS_PER_TILE(pipeline
->device
->physical_device
->rad_info
.chip_class
>= GFX9
);
1135 if (ms
->num_samples
> 1) {
1136 unsigned log_samples
= util_logbase2(ms
->num_samples
);
1137 unsigned log_ps_iter_samples
= util_logbase2(util_next_power_of_two(ps_iter_samples
));
1138 ms
->pa_sc_mode_cntl_0
|= S_028A48_MSAA_ENABLE(1);
1139 ms
->pa_sc_line_cntl
|= S_028BDC_EXPAND_LINE_WIDTH(1); /* CM_R_028BDC_PA_SC_LINE_CNTL */
1140 ms
->db_eqaa
|= S_028804_MAX_ANCHOR_SAMPLES(log_samples
) |
1141 S_028804_PS_ITER_SAMPLES(log_ps_iter_samples
) |
1142 S_028804_MASK_EXPORT_NUM_SAMPLES(log_samples
) |
1143 S_028804_ALPHA_TO_MASK_NUM_SAMPLES(log_samples
);
1144 ms
->pa_sc_aa_config
|= S_028BE0_MSAA_NUM_SAMPLES(log_samples
) |
1145 S_028BE0_MAX_SAMPLE_DIST(radv_cayman_get_maxdist(log_samples
)) |
1146 S_028BE0_MSAA_EXPOSED_SAMPLES(log_samples
); /* CM_R_028BE0_PA_SC_AA_CONFIG */
1147 ms
->pa_sc_mode_cntl_1
|= EG_S_028A4C_PS_ITER_SAMPLE(ps_iter_samples
> 1);
1151 if (vkms
->alphaToCoverageEnable
)
1152 blend
->db_alpha_to_mask
|= S_028B70_ALPHA_TO_MASK_ENABLE(1);
1154 if (vkms
->pSampleMask
)
1155 mask
= vkms
->pSampleMask
[0] & 0xffff;
1158 ms
->pa_sc_aa_mask
[0] = mask
| (mask
<< 16);
1159 ms
->pa_sc_aa_mask
[1] = mask
| (mask
<< 16);
1163 radv_prim_can_use_guardband(enum VkPrimitiveTopology topology
)
1166 case VK_PRIMITIVE_TOPOLOGY_POINT_LIST
:
1167 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST
:
1168 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP
:
1169 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY
:
1170 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY
:
1172 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST
:
1173 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP
:
1174 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN
:
1175 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY
:
1176 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY
:
1177 case VK_PRIMITIVE_TOPOLOGY_PATCH_LIST
:
1180 unreachable("unhandled primitive type");
1185 si_translate_prim(enum VkPrimitiveTopology topology
)
1188 case VK_PRIMITIVE_TOPOLOGY_POINT_LIST
:
1189 return V_008958_DI_PT_POINTLIST
;
1190 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST
:
1191 return V_008958_DI_PT_LINELIST
;
1192 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP
:
1193 return V_008958_DI_PT_LINESTRIP
;
1194 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST
:
1195 return V_008958_DI_PT_TRILIST
;
1196 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP
:
1197 return V_008958_DI_PT_TRISTRIP
;
1198 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN
:
1199 return V_008958_DI_PT_TRIFAN
;
1200 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY
:
1201 return V_008958_DI_PT_LINELIST_ADJ
;
1202 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY
:
1203 return V_008958_DI_PT_LINESTRIP_ADJ
;
1204 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY
:
1205 return V_008958_DI_PT_TRILIST_ADJ
;
1206 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY
:
1207 return V_008958_DI_PT_TRISTRIP_ADJ
;
1208 case VK_PRIMITIVE_TOPOLOGY_PATCH_LIST
:
1209 return V_008958_DI_PT_PATCH
;
1217 si_conv_gl_prim_to_gs_out(unsigned gl_prim
)
1220 case 0: /* GL_POINTS */
1221 return V_028A6C_OUTPRIM_TYPE_POINTLIST
;
1222 case 1: /* GL_LINES */
1223 case 3: /* GL_LINE_STRIP */
1224 case 0xA: /* GL_LINE_STRIP_ADJACENCY_ARB */
1225 case 0x8E7A: /* GL_ISOLINES */
1226 return V_028A6C_OUTPRIM_TYPE_LINESTRIP
;
1228 case 4: /* GL_TRIANGLES */
1229 case 0xc: /* GL_TRIANGLES_ADJACENCY_ARB */
1230 case 5: /* GL_TRIANGLE_STRIP */
1231 case 7: /* GL_QUADS */
1232 return V_028A6C_OUTPRIM_TYPE_TRISTRIP
;
1240 si_conv_prim_to_gs_out(enum VkPrimitiveTopology topology
)
1243 case VK_PRIMITIVE_TOPOLOGY_POINT_LIST
:
1244 case VK_PRIMITIVE_TOPOLOGY_PATCH_LIST
:
1245 return V_028A6C_OUTPRIM_TYPE_POINTLIST
;
1246 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST
:
1247 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP
:
1248 case VK_PRIMITIVE_TOPOLOGY_LINE_LIST_WITH_ADJACENCY
:
1249 case VK_PRIMITIVE_TOPOLOGY_LINE_STRIP_WITH_ADJACENCY
:
1250 return V_028A6C_OUTPRIM_TYPE_LINESTRIP
;
1251 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST
:
1252 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP
:
1253 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_FAN
:
1254 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_LIST_WITH_ADJACENCY
:
1255 case VK_PRIMITIVE_TOPOLOGY_TRIANGLE_STRIP_WITH_ADJACENCY
:
1256 return V_028A6C_OUTPRIM_TYPE_TRISTRIP
;
1263 static unsigned si_map_swizzle(unsigned swizzle
)
1267 return V_008F0C_SQ_SEL_Y
;
1269 return V_008F0C_SQ_SEL_Z
;
1271 return V_008F0C_SQ_SEL_W
;
1273 return V_008F0C_SQ_SEL_0
;
1275 return V_008F0C_SQ_SEL_1
;
1276 default: /* VK_SWIZZLE_X */
1277 return V_008F0C_SQ_SEL_X
;
1282 radv_pipeline_init_dynamic_state(struct radv_pipeline
*pipeline
,
1283 const VkGraphicsPipelineCreateInfo
*pCreateInfo
)
1285 radv_cmd_dirty_mask_t states
= RADV_CMD_DIRTY_DYNAMIC_ALL
;
1286 RADV_FROM_HANDLE(radv_render_pass
, pass
, pCreateInfo
->renderPass
);
1287 struct radv_subpass
*subpass
= &pass
->subpasses
[pCreateInfo
->subpass
];
1289 pipeline
->dynamic_state
= default_dynamic_state
;
1291 if (pCreateInfo
->pDynamicState
) {
1292 /* Remove all of the states that are marked as dynamic */
1293 uint32_t count
= pCreateInfo
->pDynamicState
->dynamicStateCount
;
1294 for (uint32_t s
= 0; s
< count
; s
++)
1295 states
&= ~(1 << pCreateInfo
->pDynamicState
->pDynamicStates
[s
]);
1298 struct radv_dynamic_state
*dynamic
= &pipeline
->dynamic_state
;
1300 /* Section 9.2 of the Vulkan 1.0.15 spec says:
1302 * pViewportState is [...] NULL if the pipeline
1303 * has rasterization disabled.
1305 if (!pCreateInfo
->pRasterizationState
->rasterizerDiscardEnable
) {
1306 assert(pCreateInfo
->pViewportState
);
1308 dynamic
->viewport
.count
= pCreateInfo
->pViewportState
->viewportCount
;
1309 if (states
& (1 << VK_DYNAMIC_STATE_VIEWPORT
)) {
1310 typed_memcpy(dynamic
->viewport
.viewports
,
1311 pCreateInfo
->pViewportState
->pViewports
,
1312 pCreateInfo
->pViewportState
->viewportCount
);
1315 dynamic
->scissor
.count
= pCreateInfo
->pViewportState
->scissorCount
;
1316 if (states
& (1 << VK_DYNAMIC_STATE_SCISSOR
)) {
1317 typed_memcpy(dynamic
->scissor
.scissors
,
1318 pCreateInfo
->pViewportState
->pScissors
,
1319 pCreateInfo
->pViewportState
->scissorCount
);
1323 if (states
& (1 << VK_DYNAMIC_STATE_LINE_WIDTH
)) {
1324 assert(pCreateInfo
->pRasterizationState
);
1325 dynamic
->line_width
= pCreateInfo
->pRasterizationState
->lineWidth
;
1328 if (states
& (1 << VK_DYNAMIC_STATE_DEPTH_BIAS
)) {
1329 assert(pCreateInfo
->pRasterizationState
);
1330 dynamic
->depth_bias
.bias
=
1331 pCreateInfo
->pRasterizationState
->depthBiasConstantFactor
;
1332 dynamic
->depth_bias
.clamp
=
1333 pCreateInfo
->pRasterizationState
->depthBiasClamp
;
1334 dynamic
->depth_bias
.slope
=
1335 pCreateInfo
->pRasterizationState
->depthBiasSlopeFactor
;
1338 /* Section 9.2 of the Vulkan 1.0.15 spec says:
1340 * pColorBlendState is [...] NULL if the pipeline has rasterization
1341 * disabled or if the subpass of the render pass the pipeline is
1342 * created against does not use any color attachments.
1344 bool uses_color_att
= false;
1345 for (unsigned i
= 0; i
< subpass
->color_count
; ++i
) {
1346 if (subpass
->color_attachments
[i
].attachment
!= VK_ATTACHMENT_UNUSED
) {
1347 uses_color_att
= true;
1352 if (uses_color_att
&& states
& (1 << VK_DYNAMIC_STATE_BLEND_CONSTANTS
)) {
1353 assert(pCreateInfo
->pColorBlendState
);
1354 typed_memcpy(dynamic
->blend_constants
,
1355 pCreateInfo
->pColorBlendState
->blendConstants
, 4);
1358 /* If there is no depthstencil attachment, then don't read
1359 * pDepthStencilState. The Vulkan spec states that pDepthStencilState may
1360 * be NULL in this case. Even if pDepthStencilState is non-NULL, there is
1361 * no need to override the depthstencil defaults in
1362 * radv_pipeline::dynamic_state when there is no depthstencil attachment.
1364 * Section 9.2 of the Vulkan 1.0.15 spec says:
1366 * pDepthStencilState is [...] NULL if the pipeline has rasterization
1367 * disabled or if the subpass of the render pass the pipeline is created
1368 * against does not use a depth/stencil attachment.
1370 if (!pCreateInfo
->pRasterizationState
->rasterizerDiscardEnable
&&
1371 subpass
->depth_stencil_attachment
.attachment
!= VK_ATTACHMENT_UNUSED
) {
1372 assert(pCreateInfo
->pDepthStencilState
);
1374 if (states
& (1 << VK_DYNAMIC_STATE_DEPTH_BOUNDS
)) {
1375 dynamic
->depth_bounds
.min
=
1376 pCreateInfo
->pDepthStencilState
->minDepthBounds
;
1377 dynamic
->depth_bounds
.max
=
1378 pCreateInfo
->pDepthStencilState
->maxDepthBounds
;
1381 if (states
& (1 << VK_DYNAMIC_STATE_STENCIL_COMPARE_MASK
)) {
1382 dynamic
->stencil_compare_mask
.front
=
1383 pCreateInfo
->pDepthStencilState
->front
.compareMask
;
1384 dynamic
->stencil_compare_mask
.back
=
1385 pCreateInfo
->pDepthStencilState
->back
.compareMask
;
1388 if (states
& (1 << VK_DYNAMIC_STATE_STENCIL_WRITE_MASK
)) {
1389 dynamic
->stencil_write_mask
.front
=
1390 pCreateInfo
->pDepthStencilState
->front
.writeMask
;
1391 dynamic
->stencil_write_mask
.back
=
1392 pCreateInfo
->pDepthStencilState
->back
.writeMask
;
1395 if (states
& (1 << VK_DYNAMIC_STATE_STENCIL_REFERENCE
)) {
1396 dynamic
->stencil_reference
.front
=
1397 pCreateInfo
->pDepthStencilState
->front
.reference
;
1398 dynamic
->stencil_reference
.back
=
1399 pCreateInfo
->pDepthStencilState
->back
.reference
;
1403 pipeline
->dynamic_state_mask
= states
;
1406 static struct ac_shader_variant_key
1407 radv_compute_vs_key(const VkGraphicsPipelineCreateInfo
*pCreateInfo
, bool as_es
, bool as_ls
, bool export_prim_id
)
1409 struct ac_shader_variant_key key
;
1410 const VkPipelineVertexInputStateCreateInfo
*input_state
=
1411 pCreateInfo
->pVertexInputState
;
1413 memset(&key
, 0, sizeof(key
));
1414 key
.vs
.instance_rate_inputs
= 0;
1415 key
.vs
.as_es
= as_es
;
1416 key
.vs
.as_ls
= as_ls
;
1417 key
.vs
.export_prim_id
= export_prim_id
;
1419 for (unsigned i
= 0; i
< input_state
->vertexAttributeDescriptionCount
; ++i
) {
1421 binding
= input_state
->pVertexAttributeDescriptions
[i
].binding
;
1422 if (input_state
->pVertexBindingDescriptions
[binding
].inputRate
)
1423 key
.vs
.instance_rate_inputs
|= 1u << input_state
->pVertexAttributeDescriptions
[i
].location
;
1429 calculate_gs_ring_sizes(struct radv_pipeline
*pipeline
)
1431 struct radv_device
*device
= pipeline
->device
;
1432 unsigned num_se
= device
->physical_device
->rad_info
.max_se
;
1433 unsigned wave_size
= 64;
1434 unsigned max_gs_waves
= 32 * num_se
; /* max 32 per SE on GCN */
1435 unsigned gs_vertex_reuse
= 16 * num_se
; /* GS_VERTEX_REUSE register (per SE) */
1436 unsigned alignment
= 256 * num_se
;
1437 /* The maximum size is 63.999 MB per SE. */
1438 unsigned max_size
= ((unsigned)(63.999 * 1024 * 1024) & ~255) * num_se
;
1439 struct ac_shader_variant_info
*gs_info
= &pipeline
->shaders
[MESA_SHADER_GEOMETRY
]->info
;
1440 struct ac_es_output_info
*es_info
= radv_pipeline_has_tess(pipeline
) ?
1441 &pipeline
->shaders
[MESA_SHADER_TESS_EVAL
]->info
.tes
.es_info
:
1442 &pipeline
->shaders
[MESA_SHADER_VERTEX
]->info
.vs
.es_info
;
1444 /* Calculate the minimum size. */
1445 unsigned min_esgs_ring_size
= align(es_info
->esgs_itemsize
* gs_vertex_reuse
*
1446 wave_size
, alignment
);
1447 /* These are recommended sizes, not minimum sizes. */
1448 unsigned esgs_ring_size
= max_gs_waves
* 2 * wave_size
*
1449 es_info
->esgs_itemsize
* gs_info
->gs
.vertices_in
;
1450 unsigned gsvs_ring_size
= max_gs_waves
* 2 * wave_size
*
1451 gs_info
->gs
.max_gsvs_emit_size
* 1; // no streams in VK (gs->max_gs_stream + 1);
1453 min_esgs_ring_size
= align(min_esgs_ring_size
, alignment
);
1454 esgs_ring_size
= align(esgs_ring_size
, alignment
);
1455 gsvs_ring_size
= align(gsvs_ring_size
, alignment
);
1457 pipeline
->graphics
.esgs_ring_size
= CLAMP(esgs_ring_size
, min_esgs_ring_size
, max_size
);
1458 pipeline
->graphics
.gsvs_ring_size
= MIN2(gsvs_ring_size
, max_size
);
1461 static void si_multiwave_lds_size_workaround(struct radv_device
*device
,
1464 /* SPI barrier management bug:
1465 * Make sure we have at least 4k of LDS in use to avoid the bug.
1466 * It applies to workgroup sizes of more than one wavefront.
1468 if (device
->physical_device
->rad_info
.family
== CHIP_BONAIRE
||
1469 device
->physical_device
->rad_info
.family
== CHIP_KABINI
||
1470 device
->physical_device
->rad_info
.family
== CHIP_MULLINS
)
1471 *lds_size
= MAX2(*lds_size
, 8);
1475 calculate_tess_state(struct radv_pipeline
*pipeline
,
1476 const VkGraphicsPipelineCreateInfo
*pCreateInfo
)
1478 unsigned num_tcs_input_cp
= pCreateInfo
->pTessellationState
->patchControlPoints
;
1479 unsigned num_tcs_output_cp
, num_tcs_inputs
, num_tcs_outputs
;
1480 unsigned num_tcs_patch_outputs
;
1481 unsigned input_vertex_size
, output_vertex_size
, pervertex_output_patch_size
;
1482 unsigned input_patch_size
, output_patch_size
, output_patch0_offset
;
1483 unsigned lds_size
, hardware_lds_size
;
1484 unsigned perpatch_output_offset
;
1485 unsigned num_patches
;
1486 struct radv_tessellation_state
*tess
= &pipeline
->graphics
.tess
;
1488 /* This calculates how shader inputs and outputs among VS, TCS, and TES
1489 * are laid out in LDS. */
1490 num_tcs_inputs
= util_last_bit64(pipeline
->shaders
[MESA_SHADER_VERTEX
]->info
.vs
.outputs_written
);
1492 num_tcs_outputs
= util_last_bit64(pipeline
->shaders
[MESA_SHADER_TESS_CTRL
]->info
.tcs
.outputs_written
); //tcs->outputs_written
1493 num_tcs_output_cp
= pipeline
->shaders
[MESA_SHADER_TESS_CTRL
]->info
.tcs
.tcs_vertices_out
; //TCS VERTICES OUT
1494 num_tcs_patch_outputs
= util_last_bit64(pipeline
->shaders
[MESA_SHADER_TESS_CTRL
]->info
.tcs
.patch_outputs_written
);
1496 /* Ensure that we only need one wave per SIMD so we don't need to check
1497 * resource usage. Also ensures that the number of tcs in and out
1498 * vertices per threadgroup are at most 256.
1500 input_vertex_size
= num_tcs_inputs
* 16;
1501 output_vertex_size
= num_tcs_outputs
* 16;
1503 input_patch_size
= num_tcs_input_cp
* input_vertex_size
;
1505 pervertex_output_patch_size
= num_tcs_output_cp
* output_vertex_size
;
1506 output_patch_size
= pervertex_output_patch_size
+ num_tcs_patch_outputs
* 16;
1507 /* Ensure that we only need one wave per SIMD so we don't need to check
1508 * resource usage. Also ensures that the number of tcs in and out
1509 * vertices per threadgroup are at most 256.
1511 num_patches
= 64 / MAX2(num_tcs_input_cp
, num_tcs_output_cp
) * 4;
1513 /* Make sure that the data fits in LDS. This assumes the shaders only
1514 * use LDS for the inputs and outputs.
1516 hardware_lds_size
= pipeline
->device
->physical_device
->rad_info
.chip_class
>= CIK
? 65536 : 32768;
1517 num_patches
= MIN2(num_patches
, hardware_lds_size
/ (input_patch_size
+ output_patch_size
));
1519 /* Make sure the output data fits in the offchip buffer */
1520 num_patches
= MIN2(num_patches
,
1521 (pipeline
->device
->tess_offchip_block_dw_size
* 4) /
1524 /* Not necessary for correctness, but improves performance. The
1525 * specific value is taken from the proprietary driver.
1527 num_patches
= MIN2(num_patches
, 40);
1529 /* SI bug workaround - limit LS-HS threadgroups to only one wave. */
1530 if (pipeline
->device
->physical_device
->rad_info
.chip_class
== SI
) {
1531 unsigned one_wave
= 64 / MAX2(num_tcs_input_cp
, num_tcs_output_cp
);
1532 num_patches
= MIN2(num_patches
, one_wave
);
1535 output_patch0_offset
= input_patch_size
* num_patches
;
1536 perpatch_output_offset
= output_patch0_offset
+ pervertex_output_patch_size
;
1538 lds_size
= output_patch0_offset
+ output_patch_size
* num_patches
;
1540 if (pipeline
->device
->physical_device
->rad_info
.chip_class
>= CIK
) {
1541 assert(lds_size
<= 65536);
1542 lds_size
= align(lds_size
, 512) / 512;
1544 assert(lds_size
<= 32768);
1545 lds_size
= align(lds_size
, 256) / 256;
1547 si_multiwave_lds_size_workaround(pipeline
->device
, &lds_size
);
1549 tess
->lds_size
= lds_size
;
1551 tess
->tcs_in_layout
= (input_patch_size
/ 4) |
1552 ((input_vertex_size
/ 4) << 13);
1553 tess
->tcs_out_layout
= (output_patch_size
/ 4) |
1554 ((output_vertex_size
/ 4) << 13);
1555 tess
->tcs_out_offsets
= (output_patch0_offset
/ 16) |
1556 ((perpatch_output_offset
/ 16) << 16);
1557 tess
->offchip_layout
= (pervertex_output_patch_size
* num_patches
<< 16) |
1558 (num_tcs_output_cp
<< 9) | num_patches
;
1560 tess
->ls_hs_config
= S_028B58_NUM_PATCHES(num_patches
) |
1561 S_028B58_HS_NUM_INPUT_CP(num_tcs_input_cp
) |
1562 S_028B58_HS_NUM_OUTPUT_CP(num_tcs_output_cp
);
1563 tess
->num_patches
= num_patches
;
1564 tess
->num_tcs_input_cp
= num_tcs_input_cp
;
1566 struct radv_shader_variant
*tes
= pipeline
->shaders
[MESA_SHADER_TESS_EVAL
];
1567 unsigned type
= 0, partitioning
= 0, topology
= 0, distribution_mode
= 0;
1569 switch (tes
->info
.tes
.primitive_mode
) {
1571 type
= V_028B6C_TESS_TRIANGLE
;
1574 type
= V_028B6C_TESS_QUAD
;
1577 type
= V_028B6C_TESS_ISOLINE
;
1581 switch (tes
->info
.tes
.spacing
) {
1582 case TESS_SPACING_EQUAL
:
1583 partitioning
= V_028B6C_PART_INTEGER
;
1585 case TESS_SPACING_FRACTIONAL_ODD
:
1586 partitioning
= V_028B6C_PART_FRAC_ODD
;
1588 case TESS_SPACING_FRACTIONAL_EVEN
:
1589 partitioning
= V_028B6C_PART_FRAC_EVEN
;
1595 if (tes
->info
.tes
.point_mode
)
1596 topology
= V_028B6C_OUTPUT_POINT
;
1597 else if (tes
->info
.tes
.primitive_mode
== GL_ISOLINES
)
1598 topology
= V_028B6C_OUTPUT_LINE
;
1599 else if (tes
->info
.tes
.ccw
)
1600 topology
= V_028B6C_OUTPUT_TRIANGLE_CW
;
1602 topology
= V_028B6C_OUTPUT_TRIANGLE_CCW
;
1604 if (pipeline
->device
->has_distributed_tess
) {
1605 if (pipeline
->device
->physical_device
->rad_info
.family
== CHIP_FIJI
||
1606 pipeline
->device
->physical_device
->rad_info
.family
>= CHIP_POLARIS10
)
1607 distribution_mode
= V_028B6C_DISTRIBUTION_MODE_TRAPEZOIDS
;
1609 distribution_mode
= V_028B6C_DISTRIBUTION_MODE_DONUTS
;
1611 distribution_mode
= V_028B6C_DISTRIBUTION_MODE_NO_DIST
;
1613 tess
->tf_param
= S_028B6C_TYPE(type
) |
1614 S_028B6C_PARTITIONING(partitioning
) |
1615 S_028B6C_TOPOLOGY(topology
) |
1616 S_028B6C_DISTRIBUTION_MODE(distribution_mode
);
1619 static const struct radv_prim_vertex_count prim_size_table
[] = {
1620 [V_008958_DI_PT_NONE
] = {0, 0},
1621 [V_008958_DI_PT_POINTLIST
] = {1, 1},
1622 [V_008958_DI_PT_LINELIST
] = {2, 2},
1623 [V_008958_DI_PT_LINESTRIP
] = {2, 1},
1624 [V_008958_DI_PT_TRILIST
] = {3, 3},
1625 [V_008958_DI_PT_TRIFAN
] = {3, 1},
1626 [V_008958_DI_PT_TRISTRIP
] = {3, 1},
1627 [V_008958_DI_PT_LINELIST_ADJ
] = {4, 4},
1628 [V_008958_DI_PT_LINESTRIP_ADJ
] = {4, 1},
1629 [V_008958_DI_PT_TRILIST_ADJ
] = {6, 6},
1630 [V_008958_DI_PT_TRISTRIP_ADJ
] = {6, 2},
1631 [V_008958_DI_PT_RECTLIST
] = {3, 3},
1632 [V_008958_DI_PT_LINELOOP
] = {2, 1},
1633 [V_008958_DI_PT_POLYGON
] = {3, 1},
1634 [V_008958_DI_PT_2D_TRI_STRIP
] = {0, 0},
1637 static uint32_t si_vgt_gs_mode(struct radv_shader_variant
*gs
)
1639 unsigned gs_max_vert_out
= gs
->info
.gs
.vertices_out
;
1642 if (gs_max_vert_out
<= 128) {
1643 cut_mode
= V_028A40_GS_CUT_128
;
1644 } else if (gs_max_vert_out
<= 256) {
1645 cut_mode
= V_028A40_GS_CUT_256
;
1646 } else if (gs_max_vert_out
<= 512) {
1647 cut_mode
= V_028A40_GS_CUT_512
;
1649 assert(gs_max_vert_out
<= 1024);
1650 cut_mode
= V_028A40_GS_CUT_1024
;
1653 return S_028A40_MODE(V_028A40_GS_SCENARIO_G
) |
1654 S_028A40_CUT_MODE(cut_mode
)|
1655 S_028A40_ES_WRITE_OPTIMIZE(1) |
1656 S_028A40_GS_WRITE_OPTIMIZE(1);
1659 static void calculate_vgt_gs_mode(struct radv_pipeline
*pipeline
)
1661 struct radv_shader_variant
*vs
;
1662 vs
= radv_pipeline_has_gs(pipeline
) ? pipeline
->gs_copy_shader
: (radv_pipeline_has_tess(pipeline
) ? pipeline
->shaders
[MESA_SHADER_TESS_EVAL
] : pipeline
->shaders
[MESA_SHADER_VERTEX
]);
1664 struct ac_vs_output_info
*outinfo
= &vs
->info
.vs
.outinfo
;
1666 pipeline
->graphics
.vgt_primitiveid_en
= false;
1667 pipeline
->graphics
.vgt_gs_mode
= 0;
1669 if (radv_pipeline_has_gs(pipeline
)) {
1670 pipeline
->graphics
.vgt_gs_mode
= si_vgt_gs_mode(pipeline
->shaders
[MESA_SHADER_GEOMETRY
]);
1671 } else if (outinfo
->export_prim_id
) {
1672 pipeline
->graphics
.vgt_gs_mode
= S_028A40_MODE(V_028A40_GS_SCENARIO_A
);
1673 pipeline
->graphics
.vgt_primitiveid_en
= true;
1677 static void calculate_pa_cl_vs_out_cntl(struct radv_pipeline
*pipeline
)
1679 struct radv_shader_variant
*vs
;
1680 vs
= radv_pipeline_has_gs(pipeline
) ? pipeline
->gs_copy_shader
: (radv_pipeline_has_tess(pipeline
) ? pipeline
->shaders
[MESA_SHADER_TESS_EVAL
] : pipeline
->shaders
[MESA_SHADER_VERTEX
]);
1682 struct ac_vs_output_info
*outinfo
= &vs
->info
.vs
.outinfo
;
1684 unsigned clip_dist_mask
, cull_dist_mask
, total_mask
;
1685 clip_dist_mask
= outinfo
->clip_dist_mask
;
1686 cull_dist_mask
= outinfo
->cull_dist_mask
;
1687 total_mask
= clip_dist_mask
| cull_dist_mask
;
1689 bool misc_vec_ena
= outinfo
->writes_pointsize
||
1690 outinfo
->writes_layer
||
1691 outinfo
->writes_viewport_index
;
1692 pipeline
->graphics
.pa_cl_vs_out_cntl
=
1693 S_02881C_USE_VTX_POINT_SIZE(outinfo
->writes_pointsize
) |
1694 S_02881C_USE_VTX_RENDER_TARGET_INDX(outinfo
->writes_layer
) |
1695 S_02881C_USE_VTX_VIEWPORT_INDX(outinfo
->writes_viewport_index
) |
1696 S_02881C_VS_OUT_MISC_VEC_ENA(misc_vec_ena
) |
1697 S_02881C_VS_OUT_MISC_SIDE_BUS_ENA(misc_vec_ena
) |
1698 S_02881C_VS_OUT_CCDIST0_VEC_ENA((total_mask
& 0x0f) != 0) |
1699 S_02881C_VS_OUT_CCDIST1_VEC_ENA((total_mask
& 0xf0) != 0) |
1700 cull_dist_mask
<< 8 |
1705 static uint32_t offset_to_ps_input(uint32_t offset
, bool flat_shade
)
1707 uint32_t ps_input_cntl
;
1708 if (offset
<= AC_EXP_PARAM_OFFSET_31
) {
1709 ps_input_cntl
= S_028644_OFFSET(offset
);
1711 ps_input_cntl
|= S_028644_FLAT_SHADE(1);
1713 /* The input is a DEFAULT_VAL constant. */
1714 assert(offset
>= AC_EXP_PARAM_DEFAULT_VAL_0000
&&
1715 offset
<= AC_EXP_PARAM_DEFAULT_VAL_1111
);
1716 offset
-= AC_EXP_PARAM_DEFAULT_VAL_0000
;
1717 ps_input_cntl
= S_028644_OFFSET(0x20) |
1718 S_028644_DEFAULT_VAL(offset
);
1720 return ps_input_cntl
;
1723 static void calculate_ps_inputs(struct radv_pipeline
*pipeline
)
1725 struct radv_shader_variant
*ps
, *vs
;
1726 struct ac_vs_output_info
*outinfo
;
1728 ps
= pipeline
->shaders
[MESA_SHADER_FRAGMENT
];
1729 vs
= radv_pipeline_has_gs(pipeline
) ? pipeline
->gs_copy_shader
: (radv_pipeline_has_tess(pipeline
) ? pipeline
->shaders
[MESA_SHADER_TESS_EVAL
] : pipeline
->shaders
[MESA_SHADER_VERTEX
]);
1731 outinfo
= &vs
->info
.vs
.outinfo
;
1733 unsigned ps_offset
= 0;
1735 if (ps
->info
.fs
.prim_id_input
) {
1736 unsigned vs_offset
= outinfo
->vs_output_param_offset
[VARYING_SLOT_PRIMITIVE_ID
];
1737 if (vs_offset
!= AC_EXP_PARAM_UNDEFINED
) {
1738 pipeline
->graphics
.ps_input_cntl
[ps_offset
] = offset_to_ps_input(vs_offset
, true);
1743 if (ps
->info
.fs
.layer_input
) {
1744 unsigned vs_offset
= outinfo
->vs_output_param_offset
[VARYING_SLOT_LAYER
];
1745 if (vs_offset
!= AC_EXP_PARAM_UNDEFINED
)
1746 pipeline
->graphics
.ps_input_cntl
[ps_offset
] = offset_to_ps_input(vs_offset
, true);
1748 pipeline
->graphics
.ps_input_cntl
[ps_offset
] = offset_to_ps_input(AC_EXP_PARAM_DEFAULT_VAL_0000
, true);
1752 if (ps
->info
.fs
.has_pcoord
) {
1754 val
= S_028644_PT_SPRITE_TEX(1) | S_028644_OFFSET(0x20);
1755 pipeline
->graphics
.ps_input_cntl
[ps_offset
] = val
;
1759 for (unsigned i
= 0; i
< 32 && (1u << i
) <= ps
->info
.fs
.input_mask
; ++i
) {
1762 if (!(ps
->info
.fs
.input_mask
& (1u << i
)))
1765 vs_offset
= outinfo
->vs_output_param_offset
[VARYING_SLOT_VAR0
+ i
];
1766 if (vs_offset
== AC_EXP_PARAM_UNDEFINED
) {
1767 pipeline
->graphics
.ps_input_cntl
[ps_offset
] = S_028644_OFFSET(0x20);
1772 flat_shade
= !!(ps
->info
.fs
.flat_shaded_mask
& (1u << ps_offset
));
1774 pipeline
->graphics
.ps_input_cntl
[ps_offset
] = offset_to_ps_input(vs_offset
, flat_shade
);
1778 pipeline
->graphics
.ps_input_cntl_num
= ps_offset
;
1782 radv_pipeline_init(struct radv_pipeline
*pipeline
,
1783 struct radv_device
*device
,
1784 struct radv_pipeline_cache
*cache
,
1785 const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
1786 const struct radv_graphics_pipeline_create_info
*extra
,
1787 const VkAllocationCallbacks
*alloc
)
1789 struct radv_shader_module fs_m
= {0};
1791 bool has_view_index
= false;
1793 RADV_FROM_HANDLE(radv_render_pass
, pass
, pCreateInfo
->renderPass
);
1794 struct radv_subpass
*subpass
= pass
->subpasses
+ pCreateInfo
->subpass
;
1795 if (subpass
->view_mask
)
1796 has_view_index
= true;
1798 alloc
= &device
->alloc
;
1800 pipeline
->device
= device
;
1801 pipeline
->layout
= radv_pipeline_layout_from_handle(pCreateInfo
->layout
);
1803 radv_pipeline_init_dynamic_state(pipeline
, pCreateInfo
);
1804 const VkPipelineShaderStageCreateInfo
*pStages
[MESA_SHADER_STAGES
] = { 0, };
1805 struct radv_shader_module
*modules
[MESA_SHADER_STAGES
] = { 0, };
1806 for (uint32_t i
= 0; i
< pCreateInfo
->stageCount
; i
++) {
1807 gl_shader_stage stage
= ffs(pCreateInfo
->pStages
[i
].stage
) - 1;
1808 pStages
[stage
] = &pCreateInfo
->pStages
[i
];
1809 modules
[stage
] = radv_shader_module_from_handle(pStages
[stage
]->module
);
1812 radv_pipeline_init_blend_state(pipeline
, pCreateInfo
, extra
);
1814 if (!modules
[MESA_SHADER_FRAGMENT
]) {
1816 nir_builder_init_simple_shader(&fs_b
, NULL
, MESA_SHADER_FRAGMENT
, NULL
);
1817 fs_b
.shader
->info
.name
= ralloc_strdup(fs_b
.shader
, "noop_fs");
1818 fs_m
.nir
= fs_b
.shader
;
1819 modules
[MESA_SHADER_FRAGMENT
] = &fs_m
;
1822 if (modules
[MESA_SHADER_FRAGMENT
]) {
1823 struct ac_shader_variant_key key
= {0};
1824 key
.fs
.col_format
= pipeline
->graphics
.blend
.spi_shader_col_format
;
1825 if (pCreateInfo
->pMultisampleState
&&
1826 pCreateInfo
->pMultisampleState
->rasterizationSamples
> 1)
1827 key
.fs
.multisample
= true;
1829 if (pipeline
->device
->physical_device
->rad_info
.chip_class
< VI
)
1830 radv_pipeline_compute_get_int_clamp(pCreateInfo
, &key
.fs
.is_int8
, &key
.fs
.is_int10
);
1832 const VkPipelineShaderStageCreateInfo
*stage
= pStages
[MESA_SHADER_FRAGMENT
];
1834 pipeline
->shaders
[MESA_SHADER_FRAGMENT
] =
1835 radv_pipeline_compile(pipeline
, cache
, modules
[MESA_SHADER_FRAGMENT
],
1836 stage
? stage
->pName
: "main",
1837 MESA_SHADER_FRAGMENT
,
1838 stage
? stage
->pSpecializationInfo
: NULL
,
1839 pipeline
->layout
, &key
);
1840 pipeline
->active_stages
|= mesa_to_vk_shader_stage(MESA_SHADER_FRAGMENT
);
1844 ralloc_free(fs_m
.nir
);
1846 if (modules
[MESA_SHADER_VERTEX
]) {
1849 bool export_prim_id
= false;
1850 if (modules
[MESA_SHADER_TESS_CTRL
])
1852 else if (modules
[MESA_SHADER_GEOMETRY
])
1854 else if (pipeline
->shaders
[MESA_SHADER_FRAGMENT
]->info
.fs
.prim_id_input
)
1855 export_prim_id
= true;
1856 struct ac_shader_variant_key key
= radv_compute_vs_key(pCreateInfo
, as_es
, as_ls
, export_prim_id
);
1857 key
.has_multiview_view_index
= has_view_index
;
1859 pipeline
->shaders
[MESA_SHADER_VERTEX
] =
1860 radv_pipeline_compile(pipeline
, cache
, modules
[MESA_SHADER_VERTEX
],
1861 pStages
[MESA_SHADER_VERTEX
]->pName
,
1863 pStages
[MESA_SHADER_VERTEX
]->pSpecializationInfo
,
1864 pipeline
->layout
, &key
);
1866 pipeline
->active_stages
|= mesa_to_vk_shader_stage(MESA_SHADER_VERTEX
);
1869 if (modules
[MESA_SHADER_GEOMETRY
]) {
1870 struct ac_shader_variant_key key
= radv_compute_vs_key(pCreateInfo
, false, false, false);
1871 key
.has_multiview_view_index
= has_view_index
;
1873 pipeline
->shaders
[MESA_SHADER_GEOMETRY
] =
1874 radv_pipeline_compile(pipeline
, cache
, modules
[MESA_SHADER_GEOMETRY
],
1875 pStages
[MESA_SHADER_GEOMETRY
]->pName
,
1876 MESA_SHADER_GEOMETRY
,
1877 pStages
[MESA_SHADER_GEOMETRY
]->pSpecializationInfo
,
1878 pipeline
->layout
, &key
);
1880 pipeline
->active_stages
|= mesa_to_vk_shader_stage(MESA_SHADER_GEOMETRY
);
1883 if (modules
[MESA_SHADER_TESS_EVAL
]) {
1884 assert(modules
[MESA_SHADER_TESS_CTRL
]);
1886 radv_tess_pipeline_compile(pipeline
,
1888 modules
[MESA_SHADER_TESS_CTRL
],
1889 modules
[MESA_SHADER_TESS_EVAL
],
1890 pStages
[MESA_SHADER_TESS_CTRL
]->pName
,
1891 pStages
[MESA_SHADER_TESS_EVAL
]->pName
,
1892 pStages
[MESA_SHADER_TESS_CTRL
]->pSpecializationInfo
,
1893 pStages
[MESA_SHADER_TESS_EVAL
]->pSpecializationInfo
,
1895 pCreateInfo
->pTessellationState
->patchControlPoints
,
1897 pipeline
->active_stages
|= mesa_to_vk_shader_stage(MESA_SHADER_TESS_EVAL
) |
1898 mesa_to_vk_shader_stage(MESA_SHADER_TESS_CTRL
);
1901 radv_pipeline_init_depth_stencil_state(pipeline
, pCreateInfo
, extra
);
1902 radv_pipeline_init_raster_state(pipeline
, pCreateInfo
);
1903 radv_pipeline_init_multisample_state(pipeline
, pCreateInfo
);
1904 pipeline
->graphics
.prim
= si_translate_prim(pCreateInfo
->pInputAssemblyState
->topology
);
1905 pipeline
->graphics
.can_use_guardband
= radv_prim_can_use_guardband(pCreateInfo
->pInputAssemblyState
->topology
);
1907 if (radv_pipeline_has_gs(pipeline
)) {
1908 pipeline
->graphics
.gs_out
= si_conv_gl_prim_to_gs_out(pipeline
->shaders
[MESA_SHADER_GEOMETRY
]->info
.gs
.output_prim
);
1909 pipeline
->graphics
.can_use_guardband
= pipeline
->graphics
.gs_out
== V_028A6C_OUTPRIM_TYPE_TRISTRIP
;
1911 pipeline
->graphics
.gs_out
= si_conv_prim_to_gs_out(pCreateInfo
->pInputAssemblyState
->topology
);
1913 if (extra
&& extra
->use_rectlist
) {
1914 pipeline
->graphics
.prim
= V_008958_DI_PT_RECTLIST
;
1915 pipeline
->graphics
.gs_out
= V_028A6C_OUTPRIM_TYPE_TRISTRIP
;
1916 pipeline
->graphics
.can_use_guardband
= true;
1918 pipeline
->graphics
.prim_restart_enable
= !!pCreateInfo
->pInputAssemblyState
->primitiveRestartEnable
;
1919 /* prim vertex count will need TESS changes */
1920 pipeline
->graphics
.prim_vertex_count
= prim_size_table
[pipeline
->graphics
.prim
];
1922 /* Ensure that some export memory is always allocated, for two reasons:
1924 * 1) Correctness: The hardware ignores the EXEC mask if no export
1925 * memory is allocated, so KILL and alpha test do not work correctly
1927 * 2) Performance: Every shader needs at least a NULL export, even when
1928 * it writes no color/depth output. The NULL export instruction
1929 * stalls without this setting.
1931 * Don't add this to CB_SHADER_MASK.
1933 struct radv_shader_variant
*ps
= pipeline
->shaders
[MESA_SHADER_FRAGMENT
];
1934 if (!pipeline
->graphics
.blend
.spi_shader_col_format
) {
1935 if (!ps
->info
.fs
.writes_z
&&
1936 !ps
->info
.fs
.writes_stencil
&&
1937 !ps
->info
.fs
.writes_sample_mask
)
1938 pipeline
->graphics
.blend
.spi_shader_col_format
= V_028714_SPI_SHADER_32_R
;
1942 pipeline
->graphics
.db_shader_control
= 0;
1943 if (ps
->info
.fs
.early_fragment_test
|| !ps
->info
.fs
.writes_memory
)
1944 z_order
= V_02880C_EARLY_Z_THEN_LATE_Z
;
1946 z_order
= V_02880C_LATE_Z
;
1948 pipeline
->graphics
.db_shader_control
=
1949 S_02880C_Z_EXPORT_ENABLE(ps
->info
.fs
.writes_z
) |
1950 S_02880C_STENCIL_TEST_VAL_EXPORT_ENABLE(ps
->info
.fs
.writes_stencil
) |
1951 S_02880C_KILL_ENABLE(!!ps
->info
.fs
.can_discard
) |
1952 S_02880C_MASK_EXPORT_ENABLE(ps
->info
.fs
.writes_sample_mask
) |
1953 S_02880C_Z_ORDER(z_order
) |
1954 S_02880C_DEPTH_BEFORE_SHADER(ps
->info
.fs
.early_fragment_test
) |
1955 S_02880C_EXEC_ON_HIER_FAIL(ps
->info
.fs
.writes_memory
) |
1956 S_02880C_EXEC_ON_NOOP(ps
->info
.fs
.writes_memory
);
1958 if (pipeline
->device
->physical_device
->has_rbplus
)
1959 pipeline
->graphics
.db_shader_control
|= S_02880C_DUAL_QUAD_DISABLE(1);
1961 pipeline
->graphics
.shader_z_format
=
1962 ps
->info
.fs
.writes_sample_mask
? V_028710_SPI_SHADER_32_ABGR
:
1963 ps
->info
.fs
.writes_stencil
? V_028710_SPI_SHADER_32_GR
:
1964 ps
->info
.fs
.writes_z
? V_028710_SPI_SHADER_32_R
:
1965 V_028710_SPI_SHADER_ZERO
;
1967 calculate_vgt_gs_mode(pipeline
);
1968 calculate_pa_cl_vs_out_cntl(pipeline
);
1969 calculate_ps_inputs(pipeline
);
1971 for (unsigned i
= 0; i
< MESA_SHADER_STAGES
; i
++) {
1972 if (pipeline
->shaders
[i
]) {
1973 pipeline
->need_indirect_descriptor_sets
|= pipeline
->shaders
[i
]->info
.need_indirect_descriptor_sets
;
1977 uint32_t stages
= 0;
1978 if (radv_pipeline_has_tess(pipeline
)) {
1979 stages
|= S_028B54_LS_EN(V_028B54_LS_STAGE_ON
) |
1980 S_028B54_HS_EN(1) | S_028B54_DYNAMIC_HS(1);
1982 if (radv_pipeline_has_gs(pipeline
))
1983 stages
|= S_028B54_ES_EN(V_028B54_ES_STAGE_DS
) |
1985 S_028B54_VS_EN(V_028B54_VS_STAGE_COPY_SHADER
);
1987 stages
|= S_028B54_VS_EN(V_028B54_VS_STAGE_DS
);
1989 } else if (radv_pipeline_has_gs(pipeline
))
1990 stages
|= S_028B54_ES_EN(V_028B54_ES_STAGE_REAL
) |
1992 S_028B54_VS_EN(V_028B54_VS_STAGE_COPY_SHADER
);
1994 if (device
->physical_device
->rad_info
.chip_class
>= GFX9
)
1995 stages
|= S_028B54_MAX_PRIMGRP_IN_WAVE(2);
1997 pipeline
->graphics
.vgt_shader_stages_en
= stages
;
1999 if (radv_pipeline_has_gs(pipeline
))
2000 calculate_gs_ring_sizes(pipeline
);
2002 if (radv_pipeline_has_tess(pipeline
)) {
2003 if (pipeline
->graphics
.prim
== V_008958_DI_PT_PATCH
) {
2004 pipeline
->graphics
.prim_vertex_count
.min
= pCreateInfo
->pTessellationState
->patchControlPoints
;
2005 pipeline
->graphics
.prim_vertex_count
.incr
= 1;
2007 calculate_tess_state(pipeline
, pCreateInfo
);
2010 const VkPipelineVertexInputStateCreateInfo
*vi_info
=
2011 pCreateInfo
->pVertexInputState
;
2012 struct radv_vertex_elements_info
*velems
= &pipeline
->vertex_elements
;
2014 for (uint32_t i
= 0; i
< vi_info
->vertexAttributeDescriptionCount
; i
++) {
2015 const VkVertexInputAttributeDescription
*desc
=
2016 &vi_info
->pVertexAttributeDescriptions
[i
];
2017 unsigned loc
= desc
->location
;
2018 const struct vk_format_description
*format_desc
;
2020 uint32_t num_format
, data_format
;
2021 format_desc
= vk_format_description(desc
->format
);
2022 first_non_void
= vk_format_get_first_non_void_channel(desc
->format
);
2024 num_format
= radv_translate_buffer_numformat(format_desc
, first_non_void
);
2025 data_format
= radv_translate_buffer_dataformat(format_desc
, first_non_void
);
2027 velems
->rsrc_word3
[loc
] = S_008F0C_DST_SEL_X(si_map_swizzle(format_desc
->swizzle
[0])) |
2028 S_008F0C_DST_SEL_Y(si_map_swizzle(format_desc
->swizzle
[1])) |
2029 S_008F0C_DST_SEL_Z(si_map_swizzle(format_desc
->swizzle
[2])) |
2030 S_008F0C_DST_SEL_W(si_map_swizzle(format_desc
->swizzle
[3])) |
2031 S_008F0C_NUM_FORMAT(num_format
) |
2032 S_008F0C_DATA_FORMAT(data_format
);
2033 velems
->format_size
[loc
] = format_desc
->block
.bits
/ 8;
2034 velems
->offset
[loc
] = desc
->offset
;
2035 velems
->binding
[loc
] = desc
->binding
;
2036 velems
->count
= MAX2(velems
->count
, loc
+ 1);
2039 for (uint32_t i
= 0; i
< vi_info
->vertexBindingDescriptionCount
; i
++) {
2040 const VkVertexInputBindingDescription
*desc
=
2041 &vi_info
->pVertexBindingDescriptions
[i
];
2043 pipeline
->binding_stride
[desc
->binding
] = desc
->stride
;
2046 struct ac_userdata_info
*loc
= radv_lookup_user_sgpr(pipeline
, MESA_SHADER_VERTEX
,
2047 AC_UD_VS_BASE_VERTEX_START_INSTANCE
);
2048 if (loc
->sgpr_idx
!= -1) {
2049 pipeline
->graphics
.vtx_base_sgpr
= radv_shader_stage_to_user_data_0(MESA_SHADER_VERTEX
, radv_pipeline_has_gs(pipeline
), radv_pipeline_has_tess(pipeline
));
2050 pipeline
->graphics
.vtx_base_sgpr
+= loc
->sgpr_idx
* 4;
2051 if (pipeline
->shaders
[MESA_SHADER_VERTEX
]->info
.info
.vs
.needs_draw_id
)
2052 pipeline
->graphics
.vtx_emit_num
= 3;
2054 pipeline
->graphics
.vtx_emit_num
= 2;
2056 if (device
->debug_flags
& RADV_DEBUG_DUMP_SHADER_STATS
) {
2057 radv_dump_pipeline_stats(device
, pipeline
);
2060 result
= radv_pipeline_scratch_init(device
, pipeline
);
2065 radv_graphics_pipeline_create(
2067 VkPipelineCache _cache
,
2068 const VkGraphicsPipelineCreateInfo
*pCreateInfo
,
2069 const struct radv_graphics_pipeline_create_info
*extra
,
2070 const VkAllocationCallbacks
*pAllocator
,
2071 VkPipeline
*pPipeline
)
2073 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2074 RADV_FROM_HANDLE(radv_pipeline_cache
, cache
, _cache
);
2075 struct radv_pipeline
*pipeline
;
2078 pipeline
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*pipeline
), 8,
2079 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2080 if (pipeline
== NULL
)
2081 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2083 memset(pipeline
, 0, sizeof(*pipeline
));
2084 result
= radv_pipeline_init(pipeline
, device
, cache
,
2085 pCreateInfo
, extra
, pAllocator
);
2086 if (result
!= VK_SUCCESS
) {
2087 radv_pipeline_destroy(device
, pipeline
, pAllocator
);
2091 *pPipeline
= radv_pipeline_to_handle(pipeline
);
2096 VkResult
radv_CreateGraphicsPipelines(
2098 VkPipelineCache pipelineCache
,
2100 const VkGraphicsPipelineCreateInfo
* pCreateInfos
,
2101 const VkAllocationCallbacks
* pAllocator
,
2102 VkPipeline
* pPipelines
)
2104 VkResult result
= VK_SUCCESS
;
2107 for (; i
< count
; i
++) {
2109 r
= radv_graphics_pipeline_create(_device
,
2112 NULL
, pAllocator
, &pPipelines
[i
]);
2113 if (r
!= VK_SUCCESS
) {
2115 pPipelines
[i
] = VK_NULL_HANDLE
;
2122 static VkResult
radv_compute_pipeline_create(
2124 VkPipelineCache _cache
,
2125 const VkComputePipelineCreateInfo
* pCreateInfo
,
2126 const VkAllocationCallbacks
* pAllocator
,
2127 VkPipeline
* pPipeline
)
2129 RADV_FROM_HANDLE(radv_device
, device
, _device
);
2130 RADV_FROM_HANDLE(radv_pipeline_cache
, cache
, _cache
);
2131 RADV_FROM_HANDLE(radv_shader_module
, module
, pCreateInfo
->stage
.module
);
2132 struct radv_pipeline
*pipeline
;
2135 pipeline
= vk_alloc2(&device
->alloc
, pAllocator
, sizeof(*pipeline
), 8,
2136 VK_SYSTEM_ALLOCATION_SCOPE_OBJECT
);
2137 if (pipeline
== NULL
)
2138 return vk_error(VK_ERROR_OUT_OF_HOST_MEMORY
);
2140 memset(pipeline
, 0, sizeof(*pipeline
));
2141 pipeline
->device
= device
;
2142 pipeline
->layout
= radv_pipeline_layout_from_handle(pCreateInfo
->layout
);
2144 pipeline
->shaders
[MESA_SHADER_COMPUTE
] =
2145 radv_pipeline_compile(pipeline
, cache
, module
,
2146 pCreateInfo
->stage
.pName
,
2147 MESA_SHADER_COMPUTE
,
2148 pCreateInfo
->stage
.pSpecializationInfo
,
2149 pipeline
->layout
, NULL
);
2152 pipeline
->need_indirect_descriptor_sets
|= pipeline
->shaders
[MESA_SHADER_COMPUTE
]->info
.need_indirect_descriptor_sets
;
2153 result
= radv_pipeline_scratch_init(device
, pipeline
);
2154 if (result
!= VK_SUCCESS
) {
2155 radv_pipeline_destroy(device
, pipeline
, pAllocator
);
2159 *pPipeline
= radv_pipeline_to_handle(pipeline
);
2161 if (device
->debug_flags
& RADV_DEBUG_DUMP_SHADER_STATS
) {
2162 radv_dump_pipeline_stats(device
, pipeline
);
2166 VkResult
radv_CreateComputePipelines(
2168 VkPipelineCache pipelineCache
,
2170 const VkComputePipelineCreateInfo
* pCreateInfos
,
2171 const VkAllocationCallbacks
* pAllocator
,
2172 VkPipeline
* pPipelines
)
2174 VkResult result
= VK_SUCCESS
;
2177 for (; i
< count
; i
++) {
2179 r
= radv_compute_pipeline_create(_device
, pipelineCache
,
2181 pAllocator
, &pPipelines
[i
]);
2182 if (r
!= VK_SUCCESS
) {
2184 pPipelines
[i
] = VK_NULL_HANDLE
;